chore: MVP deployed (#42)

First deploy done 2026-04-18. E2E extraction of the bank_statement_header
use case completes in 35 s against the live service, with 7 of 9 header
fields provenance-verified + text-agreement-green. closing_balance
asserts from spec §12 all pass.

Updates:
- README.md: status -> "MVP deployed"; worked example curl snippet;
  pointers to deployment runbook + spec + plan.
- AGENTS.md: status line updated with the live URL + date.
- pyproject.toml: version comment referencing the first deploy.
- docs/deployment.md: "First deploy" section filled in with times,
  field-level extraction result, plus a log of every small Docker/ops
  follow-up PR that had to land to make the first deploy healthy.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.env.example

@@ -4,11 +4,11 @@
 # the Postgres password.
 
 # --- Job store -----------------------------------------------------------
-IX_POSTGRES_URL=postgresql+asyncpg://infoxtractor:<password>@host.docker.internal:5431/infoxtractor
+IX_POSTGRES_URL=postgresql+asyncpg://infoxtractor:<password>@127.0.0.1:5431/infoxtractor
 
 # --- LLM backend ---------------------------------------------------------
-IX_OLLAMA_URL=http://host.docker.internal:11434
-IX_DEFAULT_MODEL=gpt-oss:20b
+IX_OLLAMA_URL=http://127.0.0.1:11434
+IX_DEFAULT_MODEL=qwen3:14b
 
 # --- OCR -----------------------------------------------------------------
 IX_OCR_ENGINE=surya

.gitignore

@@ -15,6 +15,7 @@ dist/
 build/
 *.log
 /tmp/
+.claude/
 # uv
 # uv.lock is committed intentionally for reproducible builds.
 

AGENTS.md

@@ -4,7 +4,7 @@ Async, on-prem, LLM-powered structured information extraction microservice. Give
 
 Designed to be used by other on-prem services (e.g. mammon) as a reliable fallback / second opinion for format-specific deterministic parsers.
 
-Status: design phase. Full reference spec at `docs/spec-core-pipeline.md`. MVP spec will live at `docs/superpowers/specs/`.
+Status: MVP deployed (2026-04-18) at `http://192.168.68.42:8994` — LAN only. Full reference spec at `docs/spec-core-pipeline.md`; MVP spec at `docs/superpowers/specs/2026-04-18-ix-mvp-design.md`; deploy runbook at `docs/deployment.md`.
 
 ## Guiding Principles
 
@@ -25,7 +25,7 @@ Status: design phase. Full reference spec at `docs/spec-core-pipeline.md`. MVP s
 - **Language**: Python 3.12, asyncio
 - **Web/REST**: FastAPI + uvicorn
 - **OCR (pluggable)**: Surya OCR first (GPU, shares RTX 3090 with Ollama / Immich ML)
-- **LLM**: Ollama at `192.168.68.42:11434`, structured outputs via JSON schema. Initial model candidate: `qwen2.5:32b` / `gpt-oss:20b`, configurable per use case
+- **LLM**: Ollama at `192.168.68.42:11434`, structured outputs via JSON schema. Initial model candidate: `qwen2.5:32b` / `qwen3:14b`, configurable per use case
 - **State**: Postgres on the shared `postgis` container (:5431), new `infoxtractor` database
 - **Deployment**: Docker, `git push server main` → post-receive rebuild (pattern from other apps)
 

Dockerfile

@@ -0,0 +1,69 @@
+# InfoXtractor container image.
+#
+# Base image ships CUDA 12.4 runtime libraries so the Surya OCR client can
+# use the RTX 3090 on the deploy host. Ubuntu 22.04 is the LTS used across
+# the home-server stack (immich-ml, monitoring) so GPU drivers line up.
+FROM nvidia/cuda:12.4.0-runtime-ubuntu22.04
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PIP_NO_CACHE_DIR=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# --- System deps --------------------------------------------------------
+#   - python3.12 via deadsnakes PPA (pinned; Ubuntu 22.04 ships 3.10 only)
+#   - libmagic1  : python-magic backend for MIME sniffing
+#   - libgl1     : libGL.so needed by Pillow/OpenCV wheels used by Surya
+#   - libglib2.0 : shared by Pillow/PyMuPDF headless rendering
+#   - curl       : post-receive hook's /healthz probe & general ops
+#   - ca-certs   : httpx TLS verification
+RUN apt-get update \
+ && apt-get install -y --no-install-recommends \
+        software-properties-common \
+        ca-certificates \
+        curl \
+        gnupg \
+ && add-apt-repository -y ppa:deadsnakes/ppa \
+ && apt-get update \
+ && apt-get install -y --no-install-recommends \
+        python3.12 \
+        python3.12-venv \
+        python3.12-dev \
+        libmagic1 \
+        libgl1 \
+        libglib2.0-0 \
+ && ln -sf /usr/bin/python3.12 /usr/local/bin/python \
+ && ln -sf /usr/bin/python3.12 /usr/local/bin/python3 \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+
+# --- uv (dependency resolver used by the project) -----------------------
+# Install via the standalone installer; avoids needing a working system pip
+# (python3.12 on Ubuntu 22.04 has no `distutils`, which breaks Ubuntu pip).
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
+ && ln -sf /root/.local/bin/uv /usr/local/bin/uv
+
+WORKDIR /app
+
+# Copy dependency manifests + README early so the heavy `uv sync` layer
+# caches whenever only application code changes. README.md is required
+# because pyproject.toml names it as the package's readme — hatchling
+# validates it exists when resolving the editable install.
+COPY pyproject.toml uv.lock .python-version README.md ./
+
+# Prod + OCR extras, no dev tooling. --frozen means "must match uv.lock";
+# CI catches drift before it reaches the image.
+RUN uv sync --frozen --no-dev --extra ocr
+
+# --- Application code ---------------------------------------------------
+COPY src src
+COPY alembic alembic
+COPY alembic.ini ./
+
+EXPOSE 8994
+
+# Migrations are idempotent (alembic upgrade head is a no-op on a current
+# DB) so running them on every start keeps the image + DB aligned without
+# an extra orchestration step.
+CMD ["sh", "-c", "uv run alembic upgrade head && uv run uvicorn ix.app:create_app --factory --host 0.0.0.0 --port 8994"]

README.md

@@ -4,10 +4,12 @@ Async, on-prem, LLM-powered structured information extraction microservice.
 
 Given a document (PDF, image, text) and a named *use case*, ix returns a structured JSON result whose shape matches the use-case schema — together with per-field provenance (OCR segment IDs, bounding boxes, cross-OCR agreement flags) that let the caller decide how much to trust each extracted value.
 
-**Status:** design phase. Implementation about to start.
+**Status:** MVP deployed. Live on the home LAN at `http://192.168.68.42:8994`.
 
 - Full reference spec: [`docs/spec-core-pipeline.md`](docs/spec-core-pipeline.md) (aspirational; MVP is a strict subset)
 - **MVP design:** [`docs/superpowers/specs/2026-04-18-ix-mvp-design.md`](docs/superpowers/specs/2026-04-18-ix-mvp-design.md)
+- **Implementation plan:** [`docs/superpowers/plans/2026-04-18-ix-mvp-implementation.md`](docs/superpowers/plans/2026-04-18-ix-mvp-implementation.md)
+- **Deployment runbook:** [`docs/deployment.md`](docs/deployment.md)
 - Agent / development notes: [`AGENTS.md`](AGENTS.md)
 
 ## Principles
@@ -15,3 +17,44 @@ Given a document (PDF, image, text) and a named *use case*, ix returns a structu
 - **On-prem always.** LLM = Ollama, OCR = local engines (Surya first). No OpenAI / Anthropic / Azure / AWS / cloud.
 - **Grounded extraction, not DB truth.** ix returns best-effort fields + provenance; the caller decides what to trust.
 - **Transport-agnostic pipeline core.** REST + Postgres-queue adapters in parallel on one job store.
+
+## Submitting a job
+
+```bash
+curl -X POST http://192.168.68.42:8994/jobs \
+  -H "Content-Type: application/json" \
+  -d '{
+    "use_case": "bank_statement_header",
+    "ix_client_id": "mammon",
+    "request_id": "some-correlation-id",
+    "context": {
+      "files": [{
+        "url": "http://paperless.local/api/documents/42/download/",
+        "headers": {"Authorization": "Token …"}
+      }],
+      "texts": ["<Paperless Tesseract OCR content>"]
+    }
+  }'
+# → {"job_id":"…","ix_id":"…","status":"pending"}
+```
+
+Poll `GET /jobs/{job_id}` until `status` is `done` or `error`. Optionally pass `callback_url` to receive a webhook on completion (one-shot, no retry; polling stays authoritative).
+
+Full REST surface + provenance response shape documented in the MVP design spec.
+
+## Running locally
+
+```bash
+uv sync --extra dev
+uv run pytest tests/unit -v                    # hermetic unit + integration suite
+IX_TEST_OLLAMA=1 uv run pytest tests/live -v    # needs LAN access to Ollama + GPU
+```
+
+## Deploying
+
+```bash
+git push server main      # rebuilds Docker image, restarts container, /healthz deploy gate
+python scripts/e2e_smoke.py   # E2E acceptance against the live service
+```
+
+See [`docs/deployment.md`](docs/deployment.md) for full runbook + rollback.

docker-compose.yml

@@ -0,0 +1,40 @@
+# InfoXtractor Docker Compose stack.
+#
+# Single service. Uses host networking so the container can reach:
+#   - Ollama at 127.0.0.1:11434
+#   - postgis at 127.0.0.1:5431 (bound to loopback only; security hardening)
+# Both services are LAN-hardened on the host and never exposed publicly,
+# so host-network access stays on-prem. This matches the `goldstein`
+# container pattern on the same server.
+#
+# The GPU reservation block matches immich-ml / the shape Docker Compose
+# expects for GPU allocation on this host.
+
+services:
+  infoxtractor:
+    build: .
+    container_name: infoxtractor
+    network_mode: host
+    restart: always
+    env_file: .env
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    volumes:
+      # Persist Surya (datalab) + HuggingFace model caches so rebuilds don't
+      # re-download ~1.5 GB of weights every time.
+      - ix_surya_cache:/root/.cache/datalab
+      - ix_hf_cache:/root/.cache/huggingface
+    labels:
+      infrastructure.web_url: "http://192.168.68.42:8994"
+      backup.enable: "true"
+      backup.type: "postgres"
+      backup.name: "infoxtractor"
+
+volumes:
+  ix_surya_cache:
+  ix_hf_cache:

docs/deployment.md

@@ -0,0 +1,153 @@
+# Deployment
+
+On-prem deploy to `192.168.68.42`. Push-to-deploy via a bare git repo + `post-receive` hook that rebuilds the Docker Compose stack. Pattern mirrors mammon and unified_messaging.
+
+## Topology
+
+```
+Mac (dev)
+  │  git push server main
+  ▼
+192.168.68.42:/home/server/Public/infoxtractor/repos.git   (bare)
+  │  post-receive → GIT_WORK_TREE=/…/app git checkout -f main
+  │                 docker compose up -d --build
+  │                 curl /healthz (60 s gate)
+  ▼
+Docker container `infoxtractor` (port 8994)
+  ├─ 127.0.0.1:11434  →  Ollama (qwen3:14b; host-network mode)
+  └─ 127.0.0.1:5431   →  postgis (database `infoxtractor`; host-network mode)
+```
+
+## One-time server setup
+
+Run **once** from the Mac. Idempotent.
+
+```bash
+export IX_POSTGRES_PASSWORD=<generate-a-strong-one>
+./scripts/setup_server.sh
+```
+
+The script:
+1. Creates `/home/server/Public/infoxtractor/repos.git` (bare) + `/home/server/Public/infoxtractor/app/` (worktree).
+2. Installs the `post-receive` hook (see `scripts/setup_server.sh` for the template).
+3. Creates the `infoxtractor` Postgres role + database on the shared `postgis` container.
+4. Writes `/home/server/Public/infoxtractor/app/.env` (mode 0600) from `.env.example` with the password substituted in.
+5. Verifies `qwen3:14b` is pulled in Ollama.
+6. Prints a hint to open UFW for port 8994 on the LAN subnet if it's missing.
+
+After the script finishes, add the deploy remote to the local repo:
+
+```bash
+git remote add server ssh://server@192.168.68.42/home/server/Public/infoxtractor/repos.git
+```
+
+## Normal deploy workflow
+
+```bash
+# after merging a feat branch into main
+git push server main
+
+# tail the server's deploy log
+ssh server@192.168.68.42 "tail -f /tmp/infoxtractor-deploy.log"
+
+# healthz gate (the post-receive hook also waits up to 60 s for this)
+curl http://192.168.68.42:8994/healthz
+
+# end-to-end smoke — this IS the real acceptance test
+python scripts/e2e_smoke.py
+```
+
+If the post-receive hook exits non-zero (healthz never reaches 200), the deploy is considered failed. The previous container keeps running (the hook swaps via `docker compose up -d --build`, which first builds the new image and only swaps if the build succeeds; if the new container fails `/healthz`, it's still up but broken). Investigate with `docker compose logs --tail 200` in `${APP_DIR}` and either fix forward or revert (see below).
+
+## Rollback
+
+Never force-push `main`. Rollbacks happen as **forward commits** via `git revert`:
+
+```bash
+git revert HEAD     # creates a revert commit for the last change
+git push forgejo main
+git push server main
+```
+
+## First deploy
+
+- **Date:** 2026-04-18
+- **Commit:** `fix/ollama-extract-json` (#36, the last of several Docker/ops follow-ups after PR #27 shipped the initial Dockerfile)
+- **`/healthz`:** all three probes (`postgres`, `ollama`, `ocr`) green. First-pass took ~7 min for the fresh container because Surya's recognition (1.34 GB) + detection (73 MB) models download from HuggingFace on first run; subsequent rebuilds reuse the named volumes declared in `docker-compose.yml` and come up in <30 s.
+- **E2E extraction:** `bank_statement_header` against `tests/fixtures/synthetic_giro.pdf` with Paperless-style texts:
+  - Pipeline completes in **35 s**.
+  - Extracted: `bank_name=DKB`, `account_iban=DE89370400440532013000`, `currency=EUR`, `opening_balance=1234.56`, `closing_balance=1450.22`, `statement_date=2026-03-31`, `statement_period_end=2026-03-31`, `statement_period_start=2026-03-01`, `account_type=null`.
+  - Provenance: 8 / 9 leaf fields have sources; 7 / 8 `provenance_verified` and `text_agreement` are True. `statement_period_start` shows up in the OCR but normalisation fails (dateutil picks a different interpretation of the cited day); to be chased in a follow-up.
+
+### Docker-ops follow-ups that landed during the first deploy
+
+All small, each merged as its own PR. In commit order after the scaffold (#27):
+
+- **#31** `fix(docker): uv via standalone installer` — Python 3.12 on Ubuntu 22.04 drops `distutils`; Ubuntu's pip needed it. Switched to the `uv` standalone installer, which has no pip dependency.
+- **#32** `fix(docker): include README.md in the uv sync COPY` — `hatchling` validates the readme file exists when resolving the editable project install.
+- **#33** `fix(compose): drop runtime: nvidia` — the deploy host's Docker daemon doesn't register a named `nvidia` runtime; `deploy.resources.devices` is sufficient and matches immich-ml.
+- **#34** `fix(deploy): network_mode: host` — `postgis` is bound to `127.0.0.1` on the host (security hardening T12). `host.docker.internal` points at the bridge gateway, not loopback, so the container couldn't reach postgis. Goldstein uses the same pattern.
+- **#35** `fix(deps): pin surya-ocr ^0.17` — earlier cu124 torch pin had forced surya to 0.14.1, which breaks our `surya.foundation` import and needs a transformers version that lacks `QuantizedCacheConfig`.
+- **#36** `fix(genai): drop Ollama format flag; extract trailing JSON` — Ollama 0.11.8 segfaults on Pydantic JSON Schemas (`$ref`, `anyOf`, `pattern`), and `format="json"` terminates reasoning models (qwen3) at `{}` because their `<think>…</think>` chain-of-thought isn't valid JSON. Omit the flag, inject the schema into the system prompt, extract the outermost `{…}` balanced block from the response.
+- **volumes** — named `ix_surya_cache` + `ix_hf_cache` mount `/root/.cache/datalab` + `/root/.cache/huggingface` so rebuilds don't re-download ~1.5 GB of model weights.
+
+Production notes:
+
+- `IX_DEFAULT_MODEL=qwen3:14b` (already pulled on the host). Spec listed `gpt-oss:20b` as a concrete example; swapped to keep the deploy on-prem without an extra `ollama pull`.
+- Torch 2.11 default cu13 wheels fall back to CPU against the host's CUDA 12.4 driver — Surya runs on CPU. Expected inference times: seconds per page. Upgrading the NVIDIA driver (or pinning a cu12-compatible torch wheel newer than 2.7) will unlock GPU with no code changes.
+
+## E2E smoke test (`scripts/e2e_smoke.py`)
+
+What it does (from the Mac):
+
+1. Checks `/healthz`.
+2. Starts a tiny HTTP server on the Mac's LAN IP serving `tests/fixtures/synthetic_giro.pdf`.
+3. Submits a `POST /jobs` with `use_case=bank_statement_header`, the fixture URL in `context.files`, and a Paperless-style OCR text in `context.texts` (to exercise the `text_agreement` cross-check).
+4. Polls `GET /jobs/{id}` every 2 s until terminal or 120 s timeout.
+5. Asserts: `status=="done"`, `bank_name` non-empty, `provenance.fields["result.closing_balance"].provenance_verified=True`, `text_agreement=True`, total elapsed `< 60s`.
+
+Non-zero exit means the deploy is not healthy. Roll back via `git revert HEAD`.
+
+## Operational checklists
+
+### After `ollama pull` on the host
+
+The `IX_DEFAULT_MODEL` env var on the server's `.env` must match something in `ollama list`. Changing the default means:
+
+1. Edit `/home/server/Public/infoxtractor/app/.env` → `IX_DEFAULT_MODEL=<new>`.
+2. `docker compose --project-directory /home/server/Public/infoxtractor/app restart`.
+3. `curl http://192.168.68.42:8994/healthz` → confirm `ollama: ok`.
+
+### If `/healthz` shows `ollama: degraded`
+
+`qwen3:14b` (or the configured default) is not pulled. On the host:
+```bash
+ssh server@192.168.68.42 "docker exec ollama ollama pull qwen3:14b"
+```
+
+### If `/healthz` shows `ocr: fail`
+
+Surya couldn't initialize (model missing, CUDA unavailable, OOM). First run can be slow — models download on first call. Check container logs:
+```bash
+ssh server@192.168.68.42 "docker logs infoxtractor --tail 200"
+```
+
+### If the container fails to start
+
+```bash
+ssh server@192.168.68.42 "tail -100 /tmp/infoxtractor-deploy.log"
+ssh server@192.168.68.42 "docker compose -f /home/server/Public/infoxtractor/app/docker-compose.yml logs --tail 200"
+```
+
+## Monitoring
+
+- Monitoring dashboard auto-discovers via the `infrastructure.web_url` label on the container: `http://192.168.68.42:8001` → "infoxtractor" card.
+- Backup opt-in via `backup.enable=true` + `backup.type=postgres` + `backup.name=infoxtractor` labels. The daily backup script picks up the `infoxtractor` Postgres database automatically.
+
+## Ports
+
+| Port | Direction | Source | Service |
+|------|-----------|--------|---------|
+| 8994/tcp | ALLOW | 192.168.68.0/24 | ix REST + healthz (LAN only; not publicly exposed) |
+
+No VPS Caddy entry; no `infrastructure.docs_url` label — this is an internal service.

pyproject.toml

@@ -1,6 +1,8 @@
 [project]
 name = "infoxtractor"
 version = "0.1.0"
+# Released 2026-04-18 with the first live deploy of the MVP. See
+# docs/deployment.md §"First deploy" for the commit + /healthz times.
 description = "Async on-prem LLM-powered structured information extraction microservice"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -31,10 +33,12 @@ dependencies = [
 
 [project.optional-dependencies]
 ocr = [
-    # Real OCR engine — pulls torch + CUDA wheels. Kept optional so CI
-    # (no GPU) can install the base package without the model deps.
-    "surya-ocr>=0.9",
-    "torch>=2.4",
+    # Real OCR engine. Kept optional so CI (no GPU) can install the base
+    # package without the model deps.
+    # surya >= 0.17 is required: the client code uses the
+    # `surya.foundation` module, which older releases don't expose.
+    "surya-ocr>=0.17,<0.18",
+    "torch>=2.7",
 ]
 dev = [
     "pytest>=8.3",
@@ -44,6 +48,11 @@ dev = [
     "mypy>=1.13",
 ]
 
+# Note: the default pypi torch ships cu13 wheels, which emit a
+# UserWarning and fall back to CPU against the deploy host's CUDA 12.4
+# driver. Surya then runs on CPU — slower but correct for MVP. A future
+# driver upgrade unlocks GPU Surya with no code changes.
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"

scripts/e2e_smoke.py

@@ -0,0 +1,210 @@
+"""End-to-end smoke test against the deployed infoxtractor service.
+
+Uploads a synthetic bank-statement fixture, polls for completion, and asserts
+the provenance flags per spec §12 E2E. Intended to run from the Mac after
+every `git push server main` as the deploy gate.
+
+Prerequisites:
+  - The service is running and reachable at --base-url (default
+    http://192.168.68.42:8994).
+  - The fixture `tests/fixtures/synthetic_giro.pdf` is present.
+  - The Mac and the server are on the same LAN (the server must be able to
+    reach the Mac to download the fixture).
+
+Exit codes:
+  0  all assertions passed within the timeout
+  1  at least one assertion failed
+  2  the job never reached a terminal state in time
+  3  the service was unreachable or returned an unexpected error
+
+Usage:
+  python scripts/e2e_smoke.py
+  python scripts/e2e_smoke.py --base-url http://localhost:8994
+"""
+
+from __future__ import annotations
+
+import argparse
+import http.server
+import json
+import socket
+import socketserver
+import sys
+import threading
+import time
+import urllib.error
+import urllib.request
+import uuid
+from pathlib import Path
+
+DEFAULT_BASE_URL = "http://192.168.68.42:8994"
+FIXTURE = Path(__file__).parent.parent / "tests" / "fixtures" / "synthetic_giro.pdf"
+TIMEOUT_SECONDS = 120
+POLL_INTERVAL_SECONDS = 2
+
+
+def find_lan_ip() -> str:
+    """Return the Mac's LAN IP that the server can reach."""
+    s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+    try:
+        # 192.168.68.42 is the server; getting the default route towards it
+        # yields the NIC with the matching subnet.
+        s.connect(("192.168.68.42", 80))
+        return s.getsockname()[0]
+    finally:
+        s.close()
+
+
+def serve_fixture_in_background(fixture: Path) -> tuple[str, threading.Event]:
+    """Serve the fixture on a temporary HTTP server; return the URL and a stop event."""
+    if not fixture.exists():
+        print(f"FIXTURE MISSING: {fixture}", file=sys.stderr)
+        sys.exit(3)
+
+    directory = fixture.parent
+    filename = fixture.name
+    lan_ip = find_lan_ip()
+
+    class Handler(http.server.SimpleHTTPRequestHandler):
+        def __init__(self, *args, **kwargs):
+            super().__init__(*args, directory=str(directory), **kwargs)
+
+        def log_message(self, format: str, *args) -> None:  # quiet
+            pass
+
+    # Pick any free port.
+    httpd = socketserver.TCPServer((lan_ip, 0), Handler)
+    port = httpd.server_address[1]
+    url = f"http://{lan_ip}:{port}/{filename}"
+    stop = threading.Event()
+
+    def _serve():
+        try:
+            while not stop.is_set():
+                httpd.handle_request()
+        finally:
+            httpd.server_close()
+
+    # Run in a thread. Use a loose timeout so handle_request returns when stop is set.
+    httpd.timeout = 0.5
+    t = threading.Thread(target=_serve, daemon=True)
+    t.start()
+    return url, stop
+
+
+def post_job(base_url: str, file_url: str, client_id: str, request_id: str) -> dict:
+    # Include a Paperless-style OCR of the fixture as context.texts so the
+    # text_agreement cross-check has something to compare against.
+    paperless_text = (
+        "DKB\n"
+        "DE89370400440532013000\n"
+        "Statement period: 01.03.2026 - 31.03.2026\n"
+        "Opening balance: 1234.56 EUR\n"
+        "Closing balance: 1450.22 EUR\n"
+        "31.03.2026\n"
+    )
+    payload = {
+        "use_case": "bank_statement_header",
+        "ix_client_id": client_id,
+        "request_id": request_id,
+        "context": {
+            "files": [file_url],
+            "texts": [paperless_text],
+        },
+    }
+    req = urllib.request.Request(
+        f"{base_url}/jobs",
+        data=json.dumps(payload).encode("utf-8"),
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=10) as resp:
+        return json.loads(resp.read().decode("utf-8"))
+
+
+def get_job(base_url: str, job_id: str) -> dict:
+    req = urllib.request.Request(f"{base_url}/jobs/{job_id}")
+    with urllib.request.urlopen(req, timeout=10) as resp:
+        return json.loads(resp.read().decode("utf-8"))
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--base-url", default=DEFAULT_BASE_URL)
+    parser.add_argument("--timeout", type=int, default=TIMEOUT_SECONDS)
+    args = parser.parse_args()
+
+    # Sanity-check the service is up.
+    try:
+        with urllib.request.urlopen(f"{args.base_url}/healthz", timeout=5) as resp:
+            health = json.loads(resp.read().decode("utf-8"))
+            print(f"healthz: {health}")
+    except urllib.error.URLError as e:
+        print(f"service unreachable: {e}", file=sys.stderr)
+        return 3
+
+    fixture_url, stop_server = serve_fixture_in_background(FIXTURE)
+    print(f"serving fixture at {fixture_url}")
+
+    try:
+        client_id = "e2e_smoke"
+        request_id = f"smoke-{uuid.uuid4().hex[:8]}"
+        submit = post_job(args.base_url, fixture_url, client_id, request_id)
+        job_id = submit["job_id"]
+        print(f"submitted job_id={job_id}")
+
+        started = time.monotonic()
+        last_status = None
+        job = None
+        while time.monotonic() - started < args.timeout:
+            job = get_job(args.base_url, job_id)
+            if job["status"] != last_status:
+                print(f"[{time.monotonic() - started:5.1f}s] status={job['status']}")
+                last_status = job["status"]
+            if job["status"] in ("done", "error"):
+                break
+            time.sleep(POLL_INTERVAL_SECONDS)
+        else:
+            print(f"FAIL: timed out after {args.timeout}s", file=sys.stderr)
+            return 2
+
+        assert job is not None
+        failed = []
+
+        if job["status"] != "done":
+            failed.append(f"status={job['status']!r} (want 'done')")
+
+        response = job.get("response") or {}
+        if response.get("error"):
+            failed.append(f"response.error={response['error']!r}")
+
+        result = (response.get("ix_result") or {}).get("result") or {}
+        bank = result.get("bank_name")
+        if not isinstance(bank, str) or not bank.strip():
+            failed.append(f"bank_name={bank!r} (want non-empty string)")
+
+        fields = (response.get("provenance") or {}).get("fields") or {}
+        closing = fields.get("result.closing_balance") or {}
+        if not closing.get("provenance_verified"):
+            failed.append(f"closing_balance.provenance_verified={closing.get('provenance_verified')!r}")
+        if closing.get("text_agreement") is not True:
+            failed.append(f"closing_balance.text_agreement={closing.get('text_agreement')!r} (Paperless-style text submitted)")
+
+        elapsed = time.monotonic() - started
+        if elapsed >= 60:
+            failed.append(f"elapsed={elapsed:.1f}s (≥ 60s; slow path)")
+
+        print(json.dumps(result, indent=2, default=str))
+
+        if failed:
+            print("\n".join(f"FAIL: {f}" for f in failed), file=sys.stderr)
+            return 1
+
+        print(f"\nPASS in {elapsed:.1f}s")
+        return 0
+    finally:
+        stop_server.set()
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/setup_server.sh

@@ -0,0 +1,127 @@
+#!/usr/bin/env bash
+# One-shot server setup for InfoXtractor. Idempotent: safe to re-run.
+#
+# Run from the Mac:
+#   IX_POSTGRES_PASSWORD=<pw> ./scripts/setup_server.sh
+#
+# What it does on 192.168.68.42:
+#   1. Creates the bare git repo `/home/server/Public/infoxtractor/repos.git` if missing.
+#   2. Writes the post-receive hook (or updates it) and makes it executable.
+#   3. Creates the Postgres role + database on the shared `postgis` container.
+#   4. Writes `/home/server/Public/infoxtractor/app/.env` (0600) from .env.example.
+#   5. Verifies `qwen3:14b` is pulled in Ollama.
+
+set -euo pipefail
+
+SERVER="${IX_SERVER:-server@192.168.68.42}"
+APP_BASE="/home/server/Public/infoxtractor"
+REPOS_GIT="${APP_BASE}/repos.git"
+APP_DIR="${APP_BASE}/app"
+DB_NAME="infoxtractor"
+DB_USER="infoxtractor"
+
+if [ -z "${IX_POSTGRES_PASSWORD:-}" ]; then
+  read -r -s -p "Postgres password for role '${DB_USER}': " IX_POSTGRES_PASSWORD
+  echo
+fi
+
+if [ -z "${IX_POSTGRES_PASSWORD}" ]; then
+  echo "IX_POSTGRES_PASSWORD is required." >&2
+  exit 1
+fi
+
+echo "==> 1/5  Ensuring bare repo + post-receive hook on ${SERVER}"
+ssh "${SERVER}" bash -s <<EOF
+set -euo pipefail
+mkdir -p "${REPOS_GIT}" "${APP_DIR}"
+if [ ! -f "${REPOS_GIT}/HEAD" ]; then
+  git init --bare "${REPOS_GIT}"
+fi
+
+cat >"${REPOS_GIT}/hooks/post-receive" <<'HOOK'
+#!/usr/bin/env bash
+set -eo pipefail
+
+APP_DIR="${APP_DIR}"
+LOG="/tmp/infoxtractor-deploy.log"
+
+echo "[\$(date -u '+%FT%TZ')] post-receive start" >> "\$LOG"
+
+mkdir -p "\$APP_DIR"
+GIT_WORK_TREE="\$APP_DIR" git --git-dir="${REPOS_GIT}" checkout -f main >> "\$LOG" 2>&1
+
+cd "\$APP_DIR"
+docker compose up -d --build >> "\$LOG" 2>&1
+
+# Deploy gate: /healthz must return 200 within 60 s.
+for i in \$(seq 1 30); do
+  if curl -fsS http://localhost:8994/healthz > /dev/null 2>&1; then
+    echo "[\$(date -u '+%FT%TZ')] healthz OK" >> "\$LOG"
+    exit 0
+  fi
+  sleep 2
+done
+
+echo "[\$(date -u '+%FT%TZ')] healthz never reached OK" >> "\$LOG"
+docker compose logs --tail 100 >> "\$LOG" 2>&1 || true
+exit 1
+HOOK
+
+chmod +x "${REPOS_GIT}/hooks/post-receive"
+EOF
+
+echo "==> 2/5  Verifying Ollama has qwen3:14b pulled"
+if ! ssh "${SERVER}" "docker exec ollama ollama list | awk '{print \$1}' | grep -qx 'qwen3:14b'"; then
+  echo "FAIL: qwen3:14b not found in Ollama. Run: ssh ${SERVER} 'docker exec ollama ollama pull qwen3:14b'" >&2
+  exit 1
+fi
+
+echo "==> 3/5  Creating Postgres role '${DB_USER}' and database '${DB_NAME}' on postgis container"
+# Idempotent via DO blocks; uses docker exec to avoid needing psql on the host.
+ssh "${SERVER}" bash -s <<EOF
+set -euo pipefail
+docker exec -i postgis psql -U postgres <<SQL
+DO \\\$\\\$
+BEGIN
+  IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = '${DB_USER}') THEN
+    CREATE ROLE ${DB_USER} LOGIN PASSWORD '${IX_POSTGRES_PASSWORD}';
+  ELSE
+    ALTER ROLE ${DB_USER} WITH PASSWORD '${IX_POSTGRES_PASSWORD}';
+  END IF;
+END
+\\\$\\\$;
+SQL
+
+if ! docker exec -i postgis psql -U postgres -tc "SELECT 1 FROM pg_database WHERE datname = '${DB_NAME}'" | grep -q 1; then
+  docker exec -i postgis createdb -U postgres -O ${DB_USER} ${DB_NAME}
+fi
+EOF
+
+echo "==> 4/5  Writing ${APP_DIR}/.env on the server"
+# Render .env from the repo's .env.example, substituting the password placeholder.
+LOCAL_ENV_CONTENT="$(
+  sed "s#<password>#${IX_POSTGRES_PASSWORD}#g" \
+      "$(dirname "$0")/../.env.example"
+)"
+# Append the IX_TEST_MODE=production for safety (fake mode stays off).
+# .env is written atomically and permissioned 0600.
+ssh "${SERVER}" "install -d -m 0755 '${APP_DIR}' && cat > '${APP_DIR}/.env' <<'ENVEOF'
+${LOCAL_ENV_CONTENT}
+ENVEOF
+chmod 0600 '${APP_DIR}/.env'"
+
+echo "==> 5/5  Checking UFW rule for port 8994 (LAN only)"
+ssh "${SERVER}" "sudo ufw status numbered | grep -F 8994" >/dev/null 2>&1 || {
+  echo "NOTE: UFW doesn't yet allow 8994. Run on the server:"
+  echo "  sudo ufw allow from 192.168.68.0/24 to any port 8994 proto tcp"
+}
+
+echo
+echo "Done."
+echo
+echo "Next steps (on the Mac):"
+echo "  git remote add server ssh://server@192.168.68.42${REPOS_GIT}"
+echo "  git push server main"
+echo "  ssh ${SERVER} 'tail -f /tmp/infoxtractor-deploy.log'"
+echo "  curl http://192.168.68.42:8994/healthz"
+echo "  python scripts/e2e_smoke.py"

src/ix/app.py

@@ -5,21 +5,132 @@ worker loop (Task 3.5) and the pg_queue listener (Task 3.6). Tests that
 don't care about the worker call ``create_app(spawn_worker=False)`` so the
 lifespan returns cleanly.
 
-The factory is parameterised (``spawn_worker``) instead of env-gated because
-pytest runs multiple app instances per session and we want the decision local
-to each call, not inferred from ``IX_*`` variables. The listener is also
-gated on ``spawn_worker`` — the listener is only useful when a worker is
-draining the queue, so the two share one flag.
+Task 4.3 fills in the production wiring:
+
+* Factories (``make_genai_client`` / ``make_ocr_client``) pick between
+  fakes (``IX_TEST_MODE=fake``) and real Ollama/Surya clients.
+* ``/healthz`` probes call ``selfcheck()`` on the active clients. In
+  ``fake`` mode they always report ok.
+* The worker's :class:`Pipeline` is built once per spawn with the real
+  chain of Steps; each call to the injected ``pipeline_factory`` returns
+  a fresh Pipeline so per-request state stays isolated.
 """
 
 from __future__ import annotations
 
-from collections.abc import AsyncIterator
+import asyncio
+from collections.abc import AsyncIterator, Callable
 from contextlib import asynccontextmanager, suppress
+from typing import Literal
 
 from fastapi import FastAPI
 
+from ix.adapters.rest.routes import Probes, get_probes
 from ix.adapters.rest.routes import router as rest_router
+from ix.config import AppConfig, get_config
+from ix.genai import make_genai_client
+from ix.genai.client import GenAIClient
+from ix.ocr import make_ocr_client
+from ix.ocr.client import OCRClient
+from ix.pipeline.genai_step import GenAIStep
+from ix.pipeline.ocr_step import OCRStep
+from ix.pipeline.pipeline import Pipeline
+from ix.pipeline.reliability_step import ReliabilityStep
+from ix.pipeline.response_handler_step import ResponseHandlerStep
+from ix.pipeline.setup_step import SetupStep
+
+
+def build_pipeline(
+    genai: GenAIClient, ocr: OCRClient, cfg: AppConfig
+) -> Pipeline:
+    """Assemble the production :class:`Pipeline` with injected clients.
+
+    Kept as a module-level helper so tests that want to exercise the
+    production wiring (without running the worker) can call it directly.
+    """
+
+    from pathlib import Path
+
+    from ix.ingestion import FetchConfig
+
+    return Pipeline(
+        steps=[
+            SetupStep(
+                tmp_dir=Path(cfg.tmp_dir),
+                fetch_config=FetchConfig(
+                    connect_timeout_s=float(cfg.file_connect_timeout_seconds),
+                    read_timeout_s=float(cfg.file_read_timeout_seconds),
+                    max_bytes=cfg.file_max_bytes,
+                ),
+            ),
+            OCRStep(ocr_client=ocr),
+            GenAIStep(genai_client=genai),
+            ReliabilityStep(),
+            ResponseHandlerStep(),
+        ]
+    )
+
+
+def _make_ollama_probe(
+    genai: GenAIClient, cfg: AppConfig
+) -> Callable[[], Literal["ok", "degraded", "fail"]]:
+    """Adapter: async ``selfcheck`` → sync callable the route expects.
+
+    Always drives the coroutine on a throwaway event loop in a separate
+    thread. This keeps the behavior identical whether the caller holds an
+    event loop (FastAPI request) or doesn't (a CLI tool), and avoids the
+    ``asyncio.run`` vs. already-running-loop footgun.
+    """
+
+    def probe() -> Literal["ok", "degraded", "fail"]:
+        if not hasattr(genai, "selfcheck"):
+            return "ok"  # fake client — nothing to probe.
+        return _run_async_sync(
+            lambda: genai.selfcheck(expected_model=cfg.default_model),  # type: ignore[attr-defined]
+            fallback="fail",
+        )
+
+    return probe
+
+
+def _make_ocr_probe(ocr: OCRClient) -> Callable[[], Literal["ok", "fail"]]:
+    def probe() -> Literal["ok", "fail"]:
+        if not hasattr(ocr, "selfcheck"):
+            return "ok"  # fake — nothing to probe.
+        return _run_async_sync(
+            lambda: ocr.selfcheck(),  # type: ignore[attr-defined]
+            fallback="fail",
+        )
+
+    return probe
+
+
+def _run_async_sync(make_coro, *, fallback: str) -> str:  # type: ignore[no-untyped-def]
+    """Run ``make_coro()`` on a fresh loop in a thread; return its result.
+
+    The thread owns its own event loop so the caller's loop (if any) keeps
+    running. Any exception collapses to ``fallback``.
+    """
+
+    import threading
+
+    result: dict[str, object] = {}
+
+    def _runner() -> None:
+        loop = asyncio.new_event_loop()
+        try:
+            result["value"] = loop.run_until_complete(make_coro())
+        except Exception as exc:  # any error collapses to fallback
+            result["error"] = exc
+        finally:
+            loop.close()
+
+    t = threading.Thread(target=_runner)
+    t.start()
+    t.join()
+    if "error" in result or "value" not in result:
+        return fallback
+    return str(result["value"])
 
 
 def create_app(*, spawn_worker: bool = True) -> FastAPI:
@@ -36,21 +147,35 @@ def create_app(*, spawn_worker: bool = True) -> FastAPI:
 
     @asynccontextmanager
     async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
+        cfg = get_config()
+
+        # Build the clients once per process. The worker's pipeline
+        # factory closes over these so every job runs through the same
+        # Ollama/Surya instance (Surya's predictors are heavy; re-loading
+        # them per job would be catastrophic).
+        genai_client = make_genai_client(cfg)
+        ocr_client = make_ocr_client(cfg)
+
+        # Override the route-level probe DI so /healthz reflects the
+        # actual clients. Tests that want canned probes can still override
+        # ``get_probes`` at the TestClient layer.
+        _app.dependency_overrides.setdefault(
+            get_probes,
+            lambda: Probes(
+                ollama=_make_ollama_probe(genai_client, cfg),
+                ocr=_make_ocr_probe(ocr_client),
+            ),
+        )
+
         worker_task = None
         listener = None
         if spawn_worker:
-            # Pipeline factory + listener wiring live in Chunk 4's
-            # production entrypoint; keeping this path best-effort lets the
-            # lifespan still start even on a box where Ollama/Surya aren't
-            # available (the listener just gives us a passive 10 s poll).
             try:
                 from ix.adapters.pg_queue.listener import (
                     PgQueueListener,
                     asyncpg_dsn_from_sqlalchemy_url,
                 )
-                from ix.config import get_config
 
-                cfg = get_config()
                 listener = PgQueueListener(
                     dsn=asyncpg_dsn_from_sqlalchemy_url(cfg.postgres_url)
                 )
@@ -59,10 +184,10 @@ def create_app(*, spawn_worker: bool = True) -> FastAPI:
                 listener = None
 
             try:
-                from ix.worker.loop import spawn_worker_task
-
-                worker_task = await spawn_worker_task(_app)
-            except ImportError:
+                worker_task = await _spawn_production_worker(
+                    cfg, genai_client, ocr_client, listener
+                )
+            except Exception:
                 worker_task = None
         try:
             yield
@@ -78,3 +203,30 @@ def create_app(*, spawn_worker: bool = True) -> FastAPI:
     app = FastAPI(lifespan=lifespan, title="infoxtractor", version="0.1.0")
     app.include_router(rest_router)
     return app
+
+
+async def _spawn_production_worker(
+    cfg: AppConfig,
+    genai: GenAIClient,
+    ocr: OCRClient,
+    listener,  # type: ignore[no-untyped-def]
+) -> asyncio.Task[None]:
+    """Spawn the background worker with a production pipeline factory."""
+
+    from ix.store.engine import get_session_factory
+    from ix.worker.loop import Worker
+
+    def pipeline_factory() -> Pipeline:
+        return build_pipeline(genai, ocr, cfg)
+
+    worker = Worker(
+        session_factory=get_session_factory(),
+        pipeline_factory=pipeline_factory,
+        poll_interval_seconds=10.0,
+        max_running_seconds=2 * cfg.pipeline_request_timeout_seconds,
+        callback_timeout_seconds=cfg.callback_timeout_seconds,
+        wait_for_work=listener.wait_for_work if listener is not None else None,
+    )
+
+    stop = asyncio.Event()
+    return asyncio.create_task(worker.run(stop))

src/ix/config.py

@@ -12,6 +12,7 @@ re-read after ``monkeypatch.setenv``. Production code never clears the cache.
 from __future__ import annotations
 
 from functools import lru_cache
+from typing import Literal
 
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
@@ -32,14 +33,17 @@ class AppConfig(BaseSettings):
     )
 
     # --- Job store ---
+    # Defaults assume the ix container runs with `network_mode: host` and
+    # reaches the shared `postgis` and `ollama` containers on loopback;
+    # spec §11 / docker-compose.yml ship that configuration.
     postgres_url: str = (
         "postgresql+asyncpg://infoxtractor:<password>"
-        "@host.docker.internal:5431/infoxtractor"
+        "@127.0.0.1:5431/infoxtractor"
     )
 
     # --- LLM backend ---
-    ollama_url: str = "http://host.docker.internal:11434"
-    default_model: str = "gpt-oss:20b"
+    ollama_url: str = "http://127.0.0.1:11434"
+    default_model: str = "qwen3:14b"
 
     # --- OCR ---
     ocr_engine: str = "surya"
@@ -62,6 +66,13 @@ class AppConfig(BaseSettings):
     # --- Observability ---
     log_level: str = "INFO"
 
+    # --- Test / wiring mode ---
+    # ``fake``: factories return FakeGenAIClient / FakeOCRClient and
+    # ``/healthz`` probes report ok. CI sets this so the Forgejo runner
+    # doesn't need access to Ollama or GPU-backed Surya. ``None`` (default)
+    # means production wiring: real OllamaClient + SuryaOCRClient.
+    test_mode: Literal["fake"] | None = None
+
 
 @lru_cache(maxsize=1)
 def get_config() -> AppConfig:

src/ix/genai/__init__.py

@@ -1,18 +1,43 @@
 """GenAI subsystem: protocol + fake client + invocation-result dataclasses.
 
-Real backends (Ollama, etc.) plug in behind :class:`GenAIClient`. The MVP
-ships only :class:`FakeGenAIClient` from this package; the real Ollama
-client lands in Chunk 4.
+Real backends (Ollama, …) plug in behind :class:`GenAIClient`. The factory
+:func:`make_genai_client` picks between :class:`FakeGenAIClient` (for CI
+/ hermetic tests via ``IX_TEST_MODE=fake``) and :class:`OllamaClient`
+(production). Tests that want a real Ollama client anyway can call the
+constructor directly.
 """
 
 from __future__ import annotations
 
+from ix.config import AppConfig
 from ix.genai.client import GenAIClient, GenAIInvocationResult, GenAIUsage
 from ix.genai.fake import FakeGenAIClient
+from ix.genai.ollama_client import OllamaClient
+
+
+def make_genai_client(cfg: AppConfig) -> GenAIClient:
+    """Return the :class:`GenAIClient` configured for the current run.
+
+    When ``cfg.test_mode == "fake"`` the fake is returned; the pipeline
+    callers are expected to override the injected client via DI if they
+    want a non-default canned response. Otherwise a live
+    :class:`OllamaClient` bound to ``cfg.ollama_url`` and the per-call
+    timeout is returned.
+    """
+
+    if cfg.test_mode == "fake":
+        return FakeGenAIClient(parsed=None)
+    return OllamaClient(
+        base_url=cfg.ollama_url,
+        per_call_timeout_s=float(cfg.genai_call_timeout_seconds),
+    )
+
 
 __all__ = [
     "FakeGenAIClient",
     "GenAIClient",
     "GenAIInvocationResult",
     "GenAIUsage",
+    "OllamaClient",
+    "make_genai_client",
 ]

src/ix/genai/ollama_client.py

@@ -36,7 +36,7 @@ class OllamaClient:
     Parameters
     ----------
     base_url:
-        Root URL of the Ollama server (e.g. ``http://host.docker.internal:11434``).
+        Root URL of the Ollama server (e.g. ``http://127.0.0.1:11434``).
         Trailing slashes are stripped.
     per_call_timeout_s:
         Hard per-call timeout for ``/api/chat``. Spec default: 1500 s.
@@ -96,8 +96,9 @@ class OllamaClient:
             ) from exc
 
         content = (payload.get("message") or {}).get("content") or ""
+        json_blob = _extract_json_blob(content)
         try:
-            parsed = response_schema.model_validate_json(content)
+            parsed = response_schema.model_validate_json(json_blob)
         except ValidationError as exc:
             raise IXException(
                 IXErrorCode.IX_002_001,
@@ -159,16 +160,39 @@ class OllamaClient:
         request_kwargs: dict[str, Any],
         response_schema: type[BaseModel],
     ) -> dict[str, Any]:
-        """Map provider-neutral kwargs to Ollama's /api/chat body."""
+        """Map provider-neutral kwargs to Ollama's /api/chat body.
+
+        Schema strategy for Ollama 0.11.8: we pass ``format="json"`` (loose
+        JSON mode) and bake the Pydantic schema into a system message
+        ahead of the caller's own system prompt. Rationale:
+
+        * The full Pydantic schema as ``format=<schema>`` crashes llama.cpp's
+          structured-output implementation (SIGSEGV) on every non-trivial
+          shape — ``anyOf`` / ``$ref`` / ``pattern`` all trigger it.
+        * ``format="json"`` alone guarantees valid JSON but not the shape;
+          models routinely return ``{}`` when not told what fields to emit.
+        * Injecting the schema into the prompt is the cheapest way to
+          get both: the model sees the expected shape explicitly, Pydantic
+          validates the response at parse time (IX_002_001 on mismatch).
+
+        Non-Ollama ``GenAIClient`` impls can ignore this behaviour and use
+        native structured-output (``response_format`` on OpenAI, etc.).
+        """
 
         messages = self._translate_messages(
             list(request_kwargs.get("messages") or [])
         )
+        messages = _inject_schema_system_message(messages, response_schema)
         body: dict[str, Any] = {
             "model": request_kwargs.get("model"),
             "messages": messages,
             "stream": False,
-            "format": response_schema.model_json_schema(),
+            # NOTE: format is deliberately omitted. `format="json"` made
+            # reasoning models (qwen3) abort after emitting `{}` because the
+            # constrained sampler terminated before the chain-of-thought
+            # finished; `format=<schema>` segfaulted Ollama 0.11.8. Letting
+            # the model stream freely and then extracting the trailing JSON
+            # blob works for both reasoning and non-reasoning models.
         }
 
         options: dict[str, Any] = {}
@@ -200,4 +224,117 @@ class OllamaClient:
         return out
 
 
+def _extract_json_blob(text: str) -> str:
+    """Return the outermost balanced JSON object in ``text``.
+
+    Reasoning models (qwen3, deepseek-r1) wrap their real answer in
+    ``<think>…</think>`` blocks. Other models sometimes prefix prose or
+    fence the JSON in ```json``` code blocks. Finding the last balanced
+    ``{…}`` is the cheapest robust parse that works for all three shapes;
+    a malformed response yields the full text and Pydantic catches it
+    downstream as ``IX_002_001``.
+    """
+    start = text.find("{")
+    if start < 0:
+        return text
+    depth = 0
+    in_string = False
+    escaped = False
+    for i in range(start, len(text)):
+        ch = text[i]
+        if in_string:
+            if escaped:
+                escaped = False
+            elif ch == "\\":
+                escaped = True
+            elif ch == '"':
+                in_string = False
+            continue
+        if ch == '"':
+            in_string = True
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return text[start : i + 1]
+    return text[start:]
+
+
+def _inject_schema_system_message(
+    messages: list[dict[str, Any]],
+    response_schema: type[BaseModel],
+) -> list[dict[str, Any]]:
+    """Prepend a system message that pins the expected JSON shape.
+
+    Ollama's ``format="json"`` mode guarantees valid JSON but not the
+    field set or names. We emit the Pydantic schema as JSON and
+    instruct the model to match it. If the caller already provides a
+    system message, we prepend ours; otherwise ours becomes the first
+    system turn.
+    """
+    import json as _json
+
+    schema_json = _json.dumps(
+        _sanitise_schema_for_ollama(response_schema.model_json_schema()),
+        indent=2,
+    )
+    guidance = (
+        "Respond ONLY with a single JSON object matching this JSON Schema "
+        "exactly. No prose, no code fences, no explanations. All top-level "
+        "properties listed in `required` MUST be present. Use null for "
+        "fields you cannot confidently extract. The JSON Schema:\n"
+        f"{schema_json}"
+    )
+    return [{"role": "system", "content": guidance}, *messages]
+
+
+def _sanitise_schema_for_ollama(schema: Any) -> Any:
+    """Strip null branches from ``anyOf`` unions.
+
+    Ollama 0.11.8's llama.cpp structured-output implementation segfaults on
+    Pydantic v2's standard Optional pattern::
+
+        {"anyOf": [{"type": "string"}, {"type": "null"}]}
+
+    We collapse any ``anyOf`` that includes a ``{"type": "null"}`` entry to
+    its non-null branch — single branch becomes that branch inline; multiple
+    branches keep the union without null. This only narrows what the LLM is
+    *told* it may emit; Pydantic still validates the real response and can
+    accept ``None`` at parse time if the field is ``Optional``.
+
+    Walk is recursive and structure-preserving. Other ``anyOf`` shapes (e.g.
+    polymorphic unions without null) are left alone.
+    """
+    if isinstance(schema, dict):
+        cleaned: dict[str, Any] = {}
+        for key, value in schema.items():
+            if key == "anyOf" and isinstance(value, list):
+                non_null = [
+                    _sanitise_schema_for_ollama(branch)
+                    for branch in value
+                    if not (isinstance(branch, dict) and branch.get("type") == "null")
+                ]
+                if len(non_null) == 1:
+                    # Inline the single remaining branch; merge its keys into the
+                    # parent so siblings like ``default``/``title`` are preserved.
+                    only = non_null[0]
+                    if isinstance(only, dict):
+                        for ok, ov in only.items():
+                            cleaned.setdefault(ok, ov)
+                    else:
+                        cleaned[key] = non_null
+                elif len(non_null) == 0:
+                    # Pathological: nothing left. Fall back to a permissive type.
+                    cleaned["type"] = "string"
+                else:
+                    cleaned[key] = non_null
+            else:
+                cleaned[key] = _sanitise_schema_for_ollama(value)
+        return cleaned
+    if isinstance(schema, list):
+        return [_sanitise_schema_for_ollama(item) for item in schema]
+    return schema
+
+
 __all__ = ["OllamaClient"]

src/ix/ocr/__init__.py

@@ -1,13 +1,34 @@
-"""OCR subsystem: protocol + fake client.
+"""OCR subsystem: protocol + fake + real Surya client + factory.
 
-Real engines (Surya, Azure DI, …) plug in behind :class:`OCRClient`. The
-MVP ships only :class:`FakeOCRClient` from this package; the real Surya
-client lands in Chunk 4.
+Real engines (Surya today, Azure DI / AWS Textract … deferred) plug in
+behind :class:`OCRClient`. The factory :func:`make_ocr_client` picks
+between :class:`FakeOCRClient` (when ``IX_TEST_MODE=fake``) and
+:class:`SuryaOCRClient` (production). Unknown engine names raise so a
+typo'd ``IX_OCR_ENGINE`` surfaces at startup, not later.
 """
 
 from __future__ import annotations
 
+from ix.config import AppConfig
+from ix.contracts.response import OCRDetails, OCRResult
 from ix.ocr.client import OCRClient
 from ix.ocr.fake import FakeOCRClient
+from ix.ocr.surya_client import SuryaOCRClient
 
-__all__ = ["FakeOCRClient", "OCRClient"]
+
+def make_ocr_client(cfg: AppConfig) -> OCRClient:
+    """Return the :class:`OCRClient` configured for the current run."""
+
+    if cfg.test_mode == "fake":
+        return FakeOCRClient(canned=OCRResult(result=OCRDetails()))
+    if cfg.ocr_engine == "surya":
+        return SuryaOCRClient()
+    raise ValueError(f"Unknown ocr_engine: {cfg.ocr_engine!r}")
+
+
+__all__ = [
+    "FakeOCRClient",
+    "OCRClient",
+    "SuryaOCRClient",
+    "make_ocr_client",
+]

src/ix/ocr/client.py

@@ -5,11 +5,19 @@ method satisfies the Protocol. :class:`~ix.pipeline.ocr_step.OCRStep`
 depends on the Protocol, not a concrete class, so swapping engines
 (``FakeOCRClient`` in tests, ``SuryaOCRClient`` in prod) stays a wiring
 change at the app factory.
+
+Per-page source location (``files`` + ``page_metadata``) flows in as
+optional kwargs: fakes ignore them; the real
+:class:`~ix.ocr.surya_client.SuryaOCRClient` uses them to render each
+page's pixels back from disk. Keeping these optional lets unit tests stay
+pages-only while production wiring (Task 4.3) plumbs through the real
+filesystem handles.
 """
 
 from __future__ import annotations
 
-from typing import Protocol, runtime_checkable
+from pathlib import Path
+from typing import Any, Protocol, runtime_checkable
 
 from ix.contracts import OCRResult, Page
 
@@ -24,8 +32,18 @@ class OCRClient(Protocol):
     per input page (in the same order).
     """
 
-    async def ocr(self, pages: list[Page]) -> OCRResult:
-        """Run OCR over the input pages; return the structured result."""
+    async def ocr(
+        self,
+        pages: list[Page],
+        *,
+        files: list[tuple[Path, str]] | None = None,
+        page_metadata: list[Any] | None = None,
+    ) -> OCRResult:
+        """Run OCR over the input pages; return the structured result.
+
+        ``files`` and ``page_metadata`` are optional for hermetic tests;
+        real engines that need to re-render from disk read them.
+        """
         ...
 
 

src/ix/ocr/fake.py

@@ -30,8 +30,17 @@ class FakeOCRClient:
         self._canned = canned
         self._raise_on_call = raise_on_call
 
-    async def ocr(self, pages: list[Page]) -> OCRResult:
-        """Return the canned result or raise the configured error."""
+    async def ocr(
+        self,
+        pages: list[Page],
+        **_kwargs: object,
+    ) -> OCRResult:
+        """Return the canned result or raise the configured error.
+
+        Accepts (and ignores) any keyword args the production Protocol may
+        carry — keeps the fake swappable for :class:`SuryaOCRClient` at
+        call sites that pass ``files`` / ``page_metadata``.
+        """
         if self._raise_on_call is not None:
             raise self._raise_on_call
         return self._canned

src/ix/ocr/surya_client.py

@@ -0,0 +1,235 @@
+"""SuryaOCRClient — real :class:`OCRClient` backed by ``surya-ocr``.
+
+Per spec §6.2: the MVP OCR engine. Runs Surya's detection + recognition
+predictors over per-page PIL images rendered from the downloaded sources
+(PDFs via PyMuPDF, images via Pillow).
+
+Design choices:
+
+* **Lazy model loading.** ``__init__`` is cheap; the heavy predictors are
+  built on first :meth:`ocr` / :meth:`selfcheck` / explicit :meth:`warm_up`.
+  This keeps FastAPI's lifespan predictable — ops can decide whether to
+  pay the load cost up front or on first request.
+* **Device is Surya's default.** CUDA on the prod box, MPS on M-series Macs.
+  We deliberately don't pin.
+* **No text-token reuse from PyMuPDF.** The cross-check against Paperless'
+  Tesseract output (ReliabilityStep's ``text_agreement``) is only meaningful
+  with a truly independent OCR pass, so we always render-and-recognize
+  even for PDFs that carry embedded text.
+
+The ``surya-ocr`` package pulls torch + heavy model deps, so it's kept
+behind the ``[ocr]`` extra. All Surya imports are deferred into
+:meth:`warm_up` so running the unit tests (which patch the predictors)
+doesn't require the package to be installed.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import contextlib
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Literal
+
+from ix.contracts import Line, OCRDetails, OCRResult, Page
+from ix.segmentation import PageMetadata
+
+if TYPE_CHECKING:  # pragma: no cover
+    from PIL import Image as PILImage
+
+
+class SuryaOCRClient:
+    """Surya-backed OCR engine.
+
+    Attributes are created lazily by :meth:`warm_up`. The unit tests inject
+    mocks directly onto ``_recognition_predictor`` / ``_detection_predictor``
+    to avoid the Surya import chain.
+    """
+
+    def __init__(self) -> None:
+        self._recognition_predictor: Any = None
+        self._detection_predictor: Any = None
+
+    def warm_up(self) -> None:
+        """Load the detection + recognition predictors. Idempotent.
+
+        Called automatically on the first :meth:`ocr` / :meth:`selfcheck`,
+        or explicitly from the app lifespan to front-load the cost.
+        """
+        if (
+            self._recognition_predictor is not None
+            and self._detection_predictor is not None
+        ):
+            return
+
+        # Deferred imports: only reachable when the optional [ocr] extra is
+        # installed. Keeping them inside the method so base-install unit
+        # tests (which patch the predictors) don't need surya on sys.path.
+        from surya.detection import DetectionPredictor  # type: ignore[import-not-found]
+        from surya.foundation import FoundationPredictor  # type: ignore[import-not-found]
+        from surya.recognition import RecognitionPredictor  # type: ignore[import-not-found]
+
+        foundation = FoundationPredictor()
+        self._recognition_predictor = RecognitionPredictor(foundation)
+        self._detection_predictor = DetectionPredictor()
+
+    async def ocr(
+        self,
+        pages: list[Page],
+        *,
+        files: list[tuple[Path, str]] | None = None,
+        page_metadata: list[Any] | None = None,
+    ) -> OCRResult:
+        """Render each input page, run Surya, translate back to contracts."""
+        self.warm_up()
+
+        images = self._render_pages(pages, files, page_metadata)
+
+        # Surya is blocking — run it off the event loop.
+        loop = asyncio.get_running_loop()
+        surya_results = await loop.run_in_executor(
+            None, self._run_recognition, images
+        )
+
+        out_pages: list[Page] = []
+        all_text_fragments: list[str] = []
+        for input_page, surya_result in zip(pages, surya_results, strict=True):
+            lines: list[Line] = []
+            for tl in getattr(surya_result, "text_lines", []) or []:
+                flat = self._flatten_polygon(getattr(tl, "polygon", None))
+                text = getattr(tl, "text", None)
+                lines.append(Line(text=text, bounding_box=flat))
+                if text:
+                    all_text_fragments.append(text)
+            out_pages.append(
+                Page(
+                    page_no=input_page.page_no,
+                    width=input_page.width,
+                    height=input_page.height,
+                    angle=input_page.angle,
+                    unit=input_page.unit,
+                    lines=lines,
+                )
+            )
+
+        details = OCRDetails(
+            text="\n".join(all_text_fragments) if all_text_fragments else None,
+            pages=out_pages,
+        )
+        return OCRResult(result=details, meta_data={"engine": "surya"})
+
+    async def selfcheck(self) -> Literal["ok", "fail"]:
+        """Run the predictors on a 1x1 image to confirm the stack works."""
+        try:
+            self.warm_up()
+        except Exception:
+            return "fail"
+
+        try:
+            from PIL import Image as PILImageRuntime
+
+            img = PILImageRuntime.new("RGB", (1, 1), color="white")
+            loop = asyncio.get_running_loop()
+            await loop.run_in_executor(None, self._run_recognition, [img])
+        except Exception:
+            return "fail"
+        return "ok"
+
+    def _run_recognition(self, images: list[PILImage.Image]) -> list[Any]:
+        """Invoke the recognition predictor. Kept tiny for threadpool offload."""
+        return list(
+            self._recognition_predictor(
+                images, det_predictor=self._detection_predictor
+            )
+        )
+
+    def _render_pages(
+        self,
+        pages: list[Page],
+        files: list[tuple[Path, str]] | None,
+        page_metadata: list[Any] | None,
+    ) -> list[PILImage.Image]:
+        """Render each input :class:`Page` to a PIL image.
+
+        We walk pages + page_metadata in lockstep so we know which source
+        file each page came from and (for PDFs) what page-index to render.
+        Text-only pages (``file_index is None``) get a blank 1x1 placeholder
+        so Surya returns an empty result and downstream code still gets one
+        entry per input page.
+        """
+        from PIL import Image as PILImageRuntime
+
+        metas: list[PageMetadata] = list(page_metadata or [])
+        file_records: list[tuple[Path, str]] = list(files or [])
+
+        # Per-file lazy PDF openers so we don't re-open across pages.
+        pdf_docs: dict[int, Any] = {}
+
+        # Per-file running page-within-file counter. For PDFs we emit one
+        # entry per PDF page in order; ``pages`` was built the same way by
+        # DocumentIngestor, so a parallel counter reconstructs the mapping.
+        per_file_cursor: dict[int, int] = {}
+
+        rendered: list[PILImage.Image] = []
+        try:
+            for idx, _page in enumerate(pages):
+                meta = metas[idx] if idx < len(metas) else PageMetadata()
+                file_index = meta.file_index
+                if file_index is None or file_index >= len(file_records):
+                    # Text-only page — placeholder image; Surya returns empty.
+                    rendered.append(
+                        PILImageRuntime.new("RGB", (1, 1), color="white")
+                    )
+                    continue
+
+                local_path, mime = file_records[file_index]
+                if mime == "application/pdf":
+                    doc = pdf_docs.get(file_index)
+                    if doc is None:
+                        import fitz  # PyMuPDF
+
+                        doc = fitz.open(str(local_path))
+                        pdf_docs[file_index] = doc
+                    pdf_page_no = per_file_cursor.get(file_index, 0)
+                    per_file_cursor[file_index] = pdf_page_no + 1
+                    pdf_page = doc.load_page(pdf_page_no)
+                    pix = pdf_page.get_pixmap(dpi=200)
+                    img = PILImageRuntime.frombytes(
+                        "RGB", (pix.width, pix.height), pix.samples
+                    )
+                    rendered.append(img)
+                elif mime in ("image/png", "image/jpeg", "image/tiff"):
+                    frame_no = per_file_cursor.get(file_index, 0)
+                    per_file_cursor[file_index] = frame_no + 1
+                    img = PILImageRuntime.open(local_path)
+                    # Handle multi-frame (TIFF) — seek to the right frame.
+                    with contextlib.suppress(EOFError):
+                        img.seek(frame_no)
+                    rendered.append(img.convert("RGB"))
+                else:  # pragma: no cover - ingestor already rejected
+                    rendered.append(
+                        PILImageRuntime.new("RGB", (1, 1), color="white")
+                    )
+        finally:
+            for doc in pdf_docs.values():
+                with contextlib.suppress(Exception):
+                    doc.close()
+        return rendered
+
+    @staticmethod
+    def _flatten_polygon(polygon: Any) -> list[float]:
+        """Flatten ``[[x1,y1],[x2,y2],[x3,y3],[x4,y4]]`` → 8-float list.
+
+        Surya emits 4 quad corners. The spec wants 8 raw-pixel coords so
+        downstream provenance normalisation can consume them directly.
+        """
+        if not polygon:
+            return []
+        flat: list[float] = []
+        for point in polygon:
+            if isinstance(point, (list, tuple)) and len(point) >= 2:
+                flat.append(float(point[0]))
+                flat.append(float(point[1]))
+        return flat
+
+
+__all__ = ["SuryaOCRClient"]

src/ix/pipeline/ocr_step.py

@@ -56,7 +56,11 @@ class OCRStep(Step):
         assert ctx is not None, "SetupStep must populate response_ix.context"
 
         pages = list(getattr(ctx, "pages", []))
-        ocr_result = await self._client.ocr(pages)
+        files = list(getattr(ctx, "files", []) or [])
+        page_metadata = list(getattr(ctx, "page_metadata", []) or [])
+        ocr_result = await self._client.ocr(
+            pages, files=files, page_metadata=page_metadata
+        )
 
         # Inject page tags around each OCR page's content so the LLM can
         # cross-reference the visual anchor without a separate prompt hack.

src/ix/use_cases/bank_statement_header.py

@@ -26,7 +26,7 @@ class Request(BaseModel):
     model_config = ConfigDict(extra="forbid")
 
     use_case_name: str = "Bank Statement Header"
-    default_model: str = "gpt-oss:20b"
+    default_model: str = "qwen3:14b"
     system_prompt: str = (
         "You extract header metadata from a single bank or credit-card statement. "
         "Return only facts that appear in the document; leave a field null if uncertain. "

src/ix/worker/loop.py

@@ -28,8 +28,6 @@ from collections.abc import Callable
 from datetime import UTC, datetime
 from typing import TYPE_CHECKING
 
-from fastapi import FastAPI
-
 from ix.contracts.response import ResponseIX
 from ix.errors import IXErrorCode, IXException
 from ix.pipeline.pipeline import Pipeline
@@ -179,17 +177,3 @@ class Worker:
             await session.commit()
 
 
-async def spawn_worker_task(app: FastAPI):  # type: ignore[no-untyped-def]
-    """Hook called from the FastAPI lifespan (Task 3.4).
-
-    This module-level async function is here so ``ix.app`` can import it
-    lazily without the app factory depending on the worker at import time.
-    Production wiring (Chunk 4) constructs a real Pipeline; for now we
-    build a no-op pipeline so the import chain completes. Tests that need
-    the worker wire their own Worker explicitly.
-    """
-
-    # NOTE: the real spawn is done by explicit test fixtures / a production
-    # wiring layer in Chunk 4. We return None so the lifespan's cleanup
-    # branch is a no-op; the app still runs REST fine without a worker.
-    return None

tests/live/test_ollama_client_live.py

@@ -5,7 +5,7 @@ Never runs in CI (Forgejo runner has no LAN access to Ollama). Run locally::
     IX_TEST_OLLAMA=1 uv run pytest tests/live/test_ollama_client_live.py -v
 
 Assumes the Ollama server at ``http://192.168.68.42:11434`` already has
-``gpt-oss:20b`` pulled.
+``qwen3:14b`` pulled.
 """
 
 from __future__ import annotations
@@ -26,7 +26,7 @@ pytestmark = [
 ]
 
 _OLLAMA_URL = "http://192.168.68.42:11434"
-_MODEL = "gpt-oss:20b"
+_MODEL = "qwen3:14b"
 
 
 async def test_structured_output_round_trip() -> None:

tests/live/test_surya_client_live.py

@@ -0,0 +1,83 @@
+"""Live test for :class:`SuryaOCRClient` — gated on ``IX_TEST_OLLAMA=1``.
+
+Downloads real Surya models (hundreds of MB) on first run. Never runs in
+CI. Exercised locally with::
+
+    IX_TEST_OLLAMA=1 uv run pytest tests/live/test_surya_client_live.py -v
+
+Note: requires the ``[ocr]`` extra — ``uv sync --extra ocr --extra dev``.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+import pytest
+
+from ix.contracts import Page
+from ix.segmentation import PageMetadata
+
+pytestmark = [
+    pytest.mark.live,
+    pytest.mark.skipif(
+        os.environ.get("IX_TEST_OLLAMA") != "1",
+        reason="live: IX_TEST_OLLAMA=1 required",
+    ),
+]
+
+
+async def test_extracts_dkb_and_iban_from_synthetic_giro() -> None:
+    """Real Surya run against ``tests/fixtures/synthetic_giro.pdf``.
+
+    Assert the flat text contains ``"DKB"`` and the IBAN without spaces.
+    """
+    from ix.ocr.surya_client import SuryaOCRClient
+
+    fixture = Path(__file__).parent.parent / "fixtures" / "synthetic_giro.pdf"
+    assert fixture.exists(), f"missing fixture: {fixture}"
+
+    # Build Pages the way DocumentIngestor would for this PDF: count pages
+    # via PyMuPDF so we pass the right number of inputs.
+    import fitz
+
+    doc = fitz.open(str(fixture))
+    try:
+        pages = [
+            Page(
+                page_no=i + 1,
+                width=float(p.rect.width),
+                height=float(p.rect.height),
+                lines=[],
+            )
+            for i, p in enumerate(doc)
+        ]
+    finally:
+        doc.close()
+
+    client = SuryaOCRClient()
+    result = await client.ocr(
+        pages,
+        files=[(fixture, "application/pdf")],
+        page_metadata=[PageMetadata(file_index=0) for _ in pages],
+    )
+
+    flat_text = result.result.text or ""
+    # Join page-level line texts if flat text missing (shape-safety).
+    if not flat_text:
+        flat_text = "\n".join(
+            line.text or ""
+            for page in result.result.pages
+            for line in page.lines
+        )
+
+    assert "DKB" in flat_text
+    assert "DE89370400440532013000" in flat_text.replace(" ", "")
+
+
+async def test_selfcheck_ok_against_real_predictors() -> None:
+    """``selfcheck()`` returns ``ok`` once Surya's predictors load."""
+    from ix.ocr.surya_client import SuryaOCRClient
+
+    client = SuryaOCRClient()
+    assert await client.selfcheck() == "ok"

tests/unit/test_app_wiring.py

@@ -0,0 +1,104 @@
+"""Tests for ``ix.app`` lifespan / probe wiring (Task 4.3).
+
+The lifespan selects fake clients when ``IX_TEST_MODE=fake`` and exposes
+their probes via the route DI hook. These tests exercise the probe
+adapter in isolation — no DB, no real Ollama/Surya.
+"""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from ix.app import _make_ocr_probe, _make_ollama_probe, build_pipeline
+from ix.config import AppConfig
+from ix.genai.fake import FakeGenAIClient
+from ix.ocr.fake import FakeOCRClient
+from ix.pipeline.genai_step import GenAIStep
+from ix.pipeline.ocr_step import OCRStep
+from ix.pipeline.pipeline import Pipeline
+from ix.pipeline.reliability_step import ReliabilityStep
+from ix.pipeline.response_handler_step import ResponseHandlerStep
+from ix.pipeline.setup_step import SetupStep
+
+
+def _cfg(**overrides: object) -> AppConfig:
+    return AppConfig(_env_file=None, **overrides)  # type: ignore[call-arg]
+
+
+class _SelfcheckOllamaClient:
+    async def invoke(self, *a: object, **kw: object) -> object:
+        raise NotImplementedError
+
+    async def selfcheck(
+        self, expected_model: str
+    ) -> Literal["ok", "degraded", "fail"]:
+        self.called_with = expected_model
+        return "ok"
+
+
+class _SelfcheckOCRClient:
+    async def ocr(self, *a: object, **kw: object) -> object:
+        raise NotImplementedError
+
+    async def selfcheck(self) -> Literal["ok", "fail"]:
+        return "ok"
+
+
+class _BrokenSelfcheckOllama:
+    async def invoke(self, *a: object, **kw: object) -> object:
+        raise NotImplementedError
+
+    async def selfcheck(
+        self, expected_model: str
+    ) -> Literal["ok", "degraded", "fail"]:
+        raise RuntimeError("boom")
+
+
+class TestOllamaProbe:
+    def test_fake_client_without_selfcheck_reports_ok(self) -> None:
+        cfg = _cfg(test_mode="fake", default_model="gpt-oss:20b")
+        probe = _make_ollama_probe(FakeGenAIClient(parsed=None), cfg)
+        assert probe() == "ok"
+
+    def test_real_selfcheck_returns_its_verdict(self) -> None:
+        cfg = _cfg(default_model="gpt-oss:20b")
+        client = _SelfcheckOllamaClient()
+        probe = _make_ollama_probe(client, cfg)  # type: ignore[arg-type]
+        assert probe() == "ok"
+        assert client.called_with == "gpt-oss:20b"
+
+    def test_selfcheck_exception_falls_back_to_fail(self) -> None:
+        cfg = _cfg(default_model="gpt-oss:20b")
+        probe = _make_ollama_probe(_BrokenSelfcheckOllama(), cfg)  # type: ignore[arg-type]
+        assert probe() == "fail"
+
+
+class TestOCRProbe:
+    def test_fake_client_without_selfcheck_reports_ok(self) -> None:
+        from ix.contracts.response import OCRDetails, OCRResult
+
+        probe = _make_ocr_probe(FakeOCRClient(canned=OCRResult(result=OCRDetails())))
+        assert probe() == "ok"
+
+    def test_real_selfcheck_returns_its_verdict(self) -> None:
+        probe = _make_ocr_probe(_SelfcheckOCRClient())  # type: ignore[arg-type]
+        assert probe() == "ok"
+
+
+class TestBuildPipeline:
+    def test_assembles_all_five_steps_in_order(self) -> None:
+        from ix.contracts.response import OCRDetails, OCRResult
+
+        genai = FakeGenAIClient(parsed=None)
+        ocr = FakeOCRClient(canned=OCRResult(result=OCRDetails()))
+        cfg = _cfg(test_mode="fake")
+        pipeline = build_pipeline(genai, ocr, cfg)
+        assert isinstance(pipeline, Pipeline)
+        steps = pipeline._steps  # type: ignore[attr-defined]
+        assert [type(s) for s in steps] == [
+            SetupStep,
+            OCRStep,
+            GenAIStep,
+            ReliabilityStep,
+            ResponseHandlerStep,
+        ]

tests/unit/test_config.py

@@ -51,10 +51,10 @@ def test_defaults_match_spec(monkeypatch: pytest.MonkeyPatch) -> None:
 
     assert cfg.postgres_url == (
         "postgresql+asyncpg://infoxtractor:<password>"
-        "@host.docker.internal:5431/infoxtractor"
+        "@127.0.0.1:5431/infoxtractor"
     )
-    assert cfg.ollama_url == "http://host.docker.internal:11434"
-    assert cfg.default_model == "gpt-oss:20b"
+    assert cfg.ollama_url == "http://127.0.0.1:11434"
+    assert cfg.default_model == "qwen3:14b"
     assert cfg.ocr_engine == "surya"
     assert cfg.tmp_dir == "/tmp/ix"
     assert cfg.pipeline_worker_concurrency == 1

tests/unit/test_factories.py

@@ -0,0 +1,60 @@
+"""Tests for the GenAI + OCR factories (Task 4.3).
+
+The factories pick between fake and real clients based on
+``IX_TEST_MODE``. CI runs with ``IX_TEST_MODE=fake``, production runs
+without — so the selection knob is the one lever between hermetic CI and
+real clients.
+"""
+
+from __future__ import annotations
+
+from ix.config import AppConfig
+from ix.genai import make_genai_client
+from ix.genai.fake import FakeGenAIClient
+from ix.genai.ollama_client import OllamaClient
+from ix.ocr import make_ocr_client
+from ix.ocr.fake import FakeOCRClient
+from ix.ocr.surya_client import SuryaOCRClient
+
+
+def _cfg(**overrides: object) -> AppConfig:
+    """Build an AppConfig without loading the repo's .env.example."""
+    return AppConfig(_env_file=None, **overrides)  # type: ignore[call-arg]
+
+
+class TestGenAIFactory:
+    def test_fake_mode_returns_fake(self) -> None:
+        cfg = _cfg(test_mode="fake")
+        client = make_genai_client(cfg)
+        assert isinstance(client, FakeGenAIClient)
+
+    def test_production_returns_ollama_with_configured_url(self) -> None:
+        cfg = _cfg(
+            test_mode=None,
+            ollama_url="http://ollama.host:11434",
+            genai_call_timeout_seconds=42,
+        )
+        client = make_genai_client(cfg)
+        assert isinstance(client, OllamaClient)
+        # Inspect the private attrs for binding correctness.
+        assert client._base_url == "http://ollama.host:11434"
+        assert client._per_call_timeout_s == 42
+
+
+class TestOCRFactory:
+    def test_fake_mode_returns_fake(self) -> None:
+        cfg = _cfg(test_mode="fake")
+        client = make_ocr_client(cfg)
+        assert isinstance(client, FakeOCRClient)
+
+    def test_production_surya_returns_surya(self) -> None:
+        cfg = _cfg(test_mode=None, ocr_engine="surya")
+        client = make_ocr_client(cfg)
+        assert isinstance(client, SuryaOCRClient)
+
+    def test_unknown_engine_raises(self) -> None:
+        cfg = _cfg(test_mode=None, ocr_engine="tesseract")
+        import pytest
+
+        with pytest.raises(ValueError, match="ocr_engine"):
+            make_ocr_client(cfg)

tests/unit/test_genai_step.py

@@ -363,8 +363,8 @@ class TestModelSelection:
         req = _make_request(include_provenance=False)
         resp = _response_with_segment_index(lines=["hello"])
         await step.process(req, resp)
-        # use-case default is gpt-oss:20b
-        assert client.request_kwargs["model"] == "gpt-oss:20b"  # type: ignore[index]
+        # use-case default is qwen3:14b
+        assert client.request_kwargs["model"] == "qwen3:14b"  # type: ignore[index]
 
 
 # ----------------------------------------------------------------------------

tests/unit/test_ollama_client.py

@@ -79,10 +79,19 @@ class TestInvokeHappyPath:
         body_json = json.loads(body)
         assert body_json["model"] == "gpt-oss:20b"
         assert body_json["stream"] is False
-        assert body_json["format"] == _Schema.model_json_schema()
+        # No `format` is sent: Ollama 0.11.8 segfaults on full schemas and
+        # aborts to `{}` with `format=json` on reasoning models. Schema is
+        # injected into the system prompt instead; we extract the trailing
+        # JSON blob from the response and validate via Pydantic.
+        assert "format" not in body_json
         assert body_json["options"]["temperature"] == 0.2
         assert "reasoning_effort" not in body_json
-        assert body_json["messages"] == [
+        # A schema-guidance system message is prepended to the caller's
+        # messages so Ollama (format=json loose mode) emits the right shape.
+        msgs = body_json["messages"]
+        assert msgs[0]["role"] == "system"
+        assert "JSON Schema" in msgs[0]["content"]
+        assert msgs[1:] == [
             {"role": "system", "content": "You extract."},
             {"role": "user", "content": "Doc body"},
         ]
@@ -116,7 +125,10 @@ class TestInvokeHappyPath:
         import json
 
         request_body = json.loads(httpx_mock.get_requests()[0].read())
-        assert request_body["messages"] == [
+        # First message is the auto-injected schema guidance; after that
+        # the caller's user message has its text parts joined.
+        assert request_body["messages"][0]["role"] == "system"
+        assert request_body["messages"][1:] == [
             {"role": "user", "content": "part-a\npart-b"}
         ]
 

tests/unit/test_surya_client.py

@@ -0,0 +1,166 @@
+"""Tests for :class:`SuryaOCRClient` — hermetic, no model download.
+
+The real Surya predictors are patched out with :class:`unittest.mock.MagicMock`
+that return trivially-shaped line objects. The tests assert the client's
+translation layer — flattening polygons, mapping text_lines → ``Line``,
+preserving ``page_no``/``width``/``height`` per input page.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from ix.contracts import Page
+from ix.ocr.surya_client import SuryaOCRClient
+from ix.segmentation import PageMetadata
+
+
+def _make_surya_line(text: str, polygon: list[list[float]]) -> SimpleNamespace:
+    """Mimic ``surya.recognition.schema.TextLine`` duck-typing-style."""
+    return SimpleNamespace(text=text, polygon=polygon, confidence=0.95)
+
+
+def _make_surya_ocr_result(lines: list[SimpleNamespace]) -> SimpleNamespace:
+    """Mimic ``surya.recognition.schema.OCRResult``."""
+    return SimpleNamespace(text_lines=lines, image_bbox=[0, 0, 100, 100])
+
+
+class TestOCRBuildsOCRResultFromMockedPredictors:
+    async def test_one_image_one_line_flatten_polygon(self, tmp_path: Path) -> None:
+        img_path = tmp_path / "a.png"
+        _write_tiny_png(img_path)
+
+        mock_line = _make_surya_line(
+            text="hello",
+            polygon=[[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]],
+        )
+        mock_predictor = MagicMock(
+            return_value=[_make_surya_ocr_result([mock_line])]
+        )
+
+        client = SuryaOCRClient()
+        # Skip the real warm_up; inject the mock directly.
+        client._recognition_predictor = mock_predictor
+        client._detection_predictor = MagicMock()
+
+        pages = [Page(page_no=1, width=100.0, height=50.0, lines=[])]
+        result = await client.ocr(
+            pages,
+            files=[(img_path, "image/png")],
+            page_metadata=[PageMetadata(file_index=0)],
+        )
+
+        assert len(result.result.pages) == 1
+        out_page = result.result.pages[0]
+        assert out_page.page_no == 1
+        assert out_page.width == 100.0
+        assert out_page.height == 50.0
+        assert len(out_page.lines) == 1
+        assert out_page.lines[0].text == "hello"
+        assert out_page.lines[0].bounding_box == [
+            1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0
+        ]
+
+    async def test_multiple_pages_preserves_order(self, tmp_path: Path) -> None:
+        img_a = tmp_path / "a.png"
+        img_b = tmp_path / "b.png"
+        _write_tiny_png(img_a)
+        _write_tiny_png(img_b)
+
+        mock_predictor = MagicMock(
+            return_value=[
+                _make_surya_ocr_result(
+                    [_make_surya_line("a-line", [[0, 0], [1, 0], [1, 1], [0, 1]])]
+                ),
+                _make_surya_ocr_result(
+                    [_make_surya_line("b-line", [[0, 0], [1, 0], [1, 1], [0, 1]])]
+                ),
+            ]
+        )
+
+        client = SuryaOCRClient()
+        client._recognition_predictor = mock_predictor
+        client._detection_predictor = MagicMock()
+
+        pages = [
+            Page(page_no=1, width=10.0, height=20.0, lines=[]),
+            Page(page_no=2, width=10.0, height=20.0, lines=[]),
+        ]
+        result = await client.ocr(
+            pages,
+            files=[(img_a, "image/png"), (img_b, "image/png")],
+            page_metadata=[
+                PageMetadata(file_index=0),
+                PageMetadata(file_index=1),
+            ],
+        )
+
+        assert [p.lines[0].text for p in result.result.pages] == ["a-line", "b-line"]
+
+    async def test_lazy_warm_up_on_first_ocr(self, tmp_path: Path) -> None:
+        img = tmp_path / "x.png"
+        _write_tiny_png(img)
+
+        client = SuryaOCRClient()
+
+        # Use patch.object on the instance's warm_up so we don't need real
+        # Surya module loading.
+        with patch.object(client, "warm_up", autospec=True) as mocked_warm_up:
+            # After warm_up is called, the predictors must be assigned.
+            def fake_warm_up(self: SuryaOCRClient) -> None:
+                self._recognition_predictor = MagicMock(
+                    return_value=[
+                        _make_surya_ocr_result(
+                            [
+                                _make_surya_line(
+                                    "hi", [[0, 0], [1, 0], [1, 1], [0, 1]]
+                                )
+                            ]
+                        )
+                    ]
+                )
+                self._detection_predictor = MagicMock()
+
+            mocked_warm_up.side_effect = lambda: fake_warm_up(client)
+
+            pages = [Page(page_no=1, width=10.0, height=10.0, lines=[])]
+            await client.ocr(
+                pages,
+                files=[(img, "image/png")],
+                page_metadata=[PageMetadata(file_index=0)],
+            )
+            mocked_warm_up.assert_called_once()
+
+
+class TestSelfcheck:
+    async def test_selfcheck_ok_with_mocked_predictors(self) -> None:
+        client = SuryaOCRClient()
+        client._recognition_predictor = MagicMock(
+            return_value=[_make_surya_ocr_result([])]
+        )
+        client._detection_predictor = MagicMock()
+        assert await client.selfcheck() == "ok"
+
+    async def test_selfcheck_fail_when_predictor_raises(self) -> None:
+        client = SuryaOCRClient()
+        client._recognition_predictor = MagicMock(
+            side_effect=RuntimeError("cuda broken")
+        )
+        client._detection_predictor = MagicMock()
+        assert await client.selfcheck() == "fail"
+
+
+def _write_tiny_png(path: Path) -> None:
+    """Write a 2x2 white PNG so PIL can open it."""
+    from PIL import Image
+
+    Image.new("RGB", (2, 2), color="white").save(path, format="PNG")
+
+
+@pytest.mark.parametrize("unused", [None])  # keep pytest happy if file ever runs alone
+def test_module_imports(unused: None) -> None:
+    assert SuryaOCRClient is not None

tests/unit/test_use_case_bank_statement_header.py

@@ -12,7 +12,7 @@ class TestRequest:
     def test_defaults(self) -> None:
         r = Request()
         assert r.use_case_name == "Bank Statement Header"
-        assert r.default_model == "gpt-oss:20b"
+        assert r.default_model == "qwen3:14b"
         # Stable substring for agent/worker tests that want to confirm the
         # prompt is what they think it is.
         assert "extract header metadata" in r.system_prompt

uv.lock

@@ -110,6 +110,79 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/db/3c/33bac158f8ab7f89b2e59426d5fe2e4f63f7ed25df84c036890172b412b5/cfgv-3.5.0-py2.py3-none-any.whl", hash = "sha256:a8dc6b26ad22ff227d2634a65cb388215ce6cc96bbcc5cfde7641ae87e8dacc0", size = 7445, upload-time = "2025-11-19T20:55:50.744Z" },
 ]
 
+[[package]]
+name = "charset-normalizer"
+version = "3.4.7"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e7/a1/67fe25fac3c7642725500a3f6cfe5821ad557c3abb11c9d20d12c7008d3e/charset_normalizer-3.4.7.tar.gz", hash = "sha256:ae89db9e5f98a11a4bf50407d4363e7b09b31e55bc117b4f7d80aab97ba009e5", size = 144271, upload-time = "2026-04-02T09:28:39.342Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0c/eb/4fc8d0a7110eb5fc9cc161723a34a8a6c200ce3b4fbf681bc86feee22308/charset_normalizer-3.4.7-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:eca9705049ad3c7345d574e3510665cb2cf844c2f2dcfe675332677f081cbd46", size = 311328, upload-time = "2026-04-02T09:26:24.331Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/e3/0fadc706008ac9d7b9b5be6dc767c05f9d3e5df51744ce4cc9605de7b9f4/charset_normalizer-3.4.7-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6178f72c5508bfc5fd446a5905e698c6212932f25bcdd4b47a757a50605a90e2", size = 208061, upload-time = "2026-04-02T09:26:25.568Z" },
+    { url = "https://files.pythonhosted.org/packages/42/f0/3dd1045c47f4a4604df85ec18ad093912ae1344ac706993aff91d38773a2/charset_normalizer-3.4.7-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:e1421b502d83040e6d7fb2fb18dff63957f720da3d77b2fbd3187ceb63755d7b", size = 229031, upload-time = "2026-04-02T09:26:26.865Z" },
+    { url = "https://files.pythonhosted.org/packages/dc/67/675a46eb016118a2fbde5a277a5d15f4f69d5f3f5f338e5ee2f8948fcf43/charset_normalizer-3.4.7-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:edac0f1ab77644605be2cbba52e6b7f630731fc42b34cb0f634be1a6eface56a", size = 225239, upload-time = "2026-04-02T09:26:28.044Z" },
+    { url = "https://files.pythonhosted.org/packages/4b/f8/d0118a2f5f23b02cd166fa385c60f9b0d4f9194f574e2b31cef350ad7223/charset_normalizer-3.4.7-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5649fd1c7bade02f320a462fdefd0b4bd3ce036065836d4f42e0de958038e116", size = 216589, upload-time = "2026-04-02T09:26:29.239Z" },
+    { url = "https://files.pythonhosted.org/packages/b1/f1/6d2b0b261b6c4ceef0fcb0d17a01cc5bc53586c2d4796fa04b5c540bc13d/charset_normalizer-3.4.7-cp312-cp312-manylinux_2_31_armv7l.whl", hash = "sha256:203104ed3e428044fd943bc4bf45fa73c0730391f9621e37fe39ecf477b128cb", size = 202733, upload-time = "2026-04-02T09:26:30.5Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/c0/7b1f943f7e87cc3db9626ba17807d042c38645f0a1d4415c7a14afb5591f/charset_normalizer-3.4.7-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:298930cec56029e05497a76988377cbd7457ba864beeea92ad7e844fe74cd1f1", size = 212652, upload-time = "2026-04-02T09:26:31.709Z" },
+    { url = "https://files.pythonhosted.org/packages/38/dd/5a9ab159fe45c6e72079398f277b7d2b523e7f716acc489726115a910097/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:708838739abf24b2ceb208d0e22403dd018faeef86ddac04319a62ae884c4f15", size = 211229, upload-time = "2026-04-02T09:26:33.282Z" },
+    { url = "https://files.pythonhosted.org/packages/d5/ff/531a1cad5ca855d1c1a8b69cb71abfd6d85c0291580146fda7c82857caa1/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:0f7eb884681e3938906ed0434f20c63046eacd0111c4ba96f27b76084cd679f5", size = 203552, upload-time = "2026-04-02T09:26:34.845Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/4c/a5fb52d528a8ca41f7598cb619409ece30a169fbdf9cdce592e53b46c3a6/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:4dc1e73c36828f982bfe79fadf5919923f8a6f4df2860804db9a98c48824ce8d", size = 230806, upload-time = "2026-04-02T09:26:36.152Z" },
+    { url = "https://files.pythonhosted.org/packages/59/7a/071feed8124111a32b316b33ae4de83d36923039ef8cf48120266844285b/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:aed52fea0513bac0ccde438c188c8a471c4e0f457c2dd20cdbf6ea7a450046c7", size = 212316, upload-time = "2026-04-02T09:26:37.672Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/35/f7dba3994312d7ba508e041eaac39a36b120f32d4c8662b8814dab876431/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:fea24543955a6a729c45a73fe90e08c743f0b3334bbf3201e6c4bc1b0c7fa464", size = 227274, upload-time = "2026-04-02T09:26:38.93Z" },
+    { url = "https://files.pythonhosted.org/packages/8a/2d/a572df5c9204ab7688ec1edc895a73ebded3b023bb07364710b05dd1c9be/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:bb6d88045545b26da47aa879dd4a89a71d1dce0f0e549b1abcb31dfe4a8eac49", size = 218468, upload-time = "2026-04-02T09:26:40.17Z" },
+    { url = "https://files.pythonhosted.org/packages/86/eb/890922a8b03a568ca2f336c36585a4713c55d4d67bf0f0c78924be6315ca/charset_normalizer-3.4.7-cp312-cp312-win32.whl", hash = "sha256:2257141f39fe65a3fdf38aeccae4b953e5f3b3324f4ff0daf9f15b8518666a2c", size = 148460, upload-time = "2026-04-02T09:26:41.416Z" },
+    { url = "https://files.pythonhosted.org/packages/35/d9/0e7dffa06c5ab081f75b1b786f0aefc88365825dfcd0ac544bdb7b2b6853/charset_normalizer-3.4.7-cp312-cp312-win_amd64.whl", hash = "sha256:5ed6ab538499c8644b8a3e18debabcd7ce684f3fa91cf867521a7a0279cab2d6", size = 159330, upload-time = "2026-04-02T09:26:42.554Z" },
+    { url = "https://files.pythonhosted.org/packages/9e/5d/481bcc2a7c88ea6b0878c299547843b2521ccbc40980cb406267088bc701/charset_normalizer-3.4.7-cp312-cp312-win_arm64.whl", hash = "sha256:56be790f86bfb2c98fb742ce566dfb4816e5a83384616ab59c49e0604d49c51d", size = 147828, upload-time = "2026-04-02T09:26:44.075Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/3b/66777e39d3ae1ddc77ee606be4ec6d8cbd4c801f65e5a1b6f2b11b8346dd/charset_normalizer-3.4.7-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:f496c9c3cc02230093d8330875c4c3cdfc3b73612a5fd921c65d39cbcef08063", size = 309627, upload-time = "2026-04-02T09:26:45.198Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/4e/b7f84e617b4854ade48a1b7915c8ccfadeba444d2a18c291f696e37f0d3b/charset_normalizer-3.4.7-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0ea948db76d31190bf08bd371623927ee1339d5f2a0b4b1b4a4439a65298703c", size = 207008, upload-time = "2026-04-02T09:26:46.824Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/bb/ec73c0257c9e11b268f018f068f5d00aa0ef8c8b09f7753ebd5f2880e248/charset_normalizer-3.4.7-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:a277ab8928b9f299723bc1a2dabb1265911b1a76341f90a510368ca44ad9ab66", size = 228303, upload-time = "2026-04-02T09:26:48.397Z" },
+    { url = "https://files.pythonhosted.org/packages/85/fb/32d1f5033484494619f701e719429c69b766bfc4dbc61aa9e9c8c166528b/charset_normalizer-3.4.7-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:3bec022aec2c514d9cf199522a802bd007cd588ab17ab2525f20f9c34d067c18", size = 224282, upload-time = "2026-04-02T09:26:49.684Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/07/330e3a0dda4c404d6da83b327270906e9654a24f6c546dc886a0eb0ffb23/charset_normalizer-3.4.7-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e044c39e41b92c845bc815e5ae4230804e8e7bc29e399b0437d64222d92809dd", size = 215595, upload-time = "2026-04-02T09:26:50.915Z" },
+    { url = "https://files.pythonhosted.org/packages/e3/7c/fc890655786e423f02556e0216d4b8c6bcb6bdfa890160dc66bf52dee468/charset_normalizer-3.4.7-cp313-cp313-manylinux_2_31_armv7l.whl", hash = "sha256:f495a1652cf3fbab2eb0639776dad966c2fb874d79d87ca07f9d5f059b8bd215", size = 201986, upload-time = "2026-04-02T09:26:52.197Z" },
+    { url = "https://files.pythonhosted.org/packages/d8/97/bfb18b3db2aed3b90cf54dc292ad79fdd5ad65c4eae454099475cbeadd0d/charset_normalizer-3.4.7-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:e712b419df8ba5e42b226c510472b37bd57b38e897d3eca5e8cfd410a29fa859", size = 211711, upload-time = "2026-04-02T09:26:53.49Z" },
+    { url = "https://files.pythonhosted.org/packages/6f/a5/a581c13798546a7fd557c82614a5c65a13df2157e9ad6373166d2a3e645d/charset_normalizer-3.4.7-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:7804338df6fcc08105c7745f1502ba68d900f45fd770d5bdd5288ddccb8a42d8", size = 210036, upload-time = "2026-04-02T09:26:54.975Z" },
+    { url = "https://files.pythonhosted.org/packages/8c/bf/b3ab5bcb478e4193d517644b0fb2bf5497fbceeaa7a1bc0f4d5b50953861/charset_normalizer-3.4.7-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:481551899c856c704d58119b5025793fa6730adda3571971af568f66d2424bb5", size = 202998, upload-time = "2026-04-02T09:26:56.303Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/4e/23efd79b65d314fa320ec6017b4b5834d5c12a58ba4610aa353af2e2f577/charset_normalizer-3.4.7-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:f59099f9b66f0d7145115e6f80dd8b1d847176df89b234a5a6b3f00437aa0832", size = 230056, upload-time = "2026-04-02T09:26:57.554Z" },
+    { url = "https://files.pythonhosted.org/packages/b9/9f/1e1941bc3f0e01df116e68dc37a55c4d249df5e6fa77f008841aef68264f/charset_normalizer-3.4.7-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:f59ad4c0e8f6bba240a9bb85504faa1ab438237199d4cce5f622761507b8f6a6", size = 211537, upload-time = "2026-04-02T09:26:58.843Z" },
+    { url = "https://files.pythonhosted.org/packages/80/0f/088cbb3020d44428964a6c97fe1edfb1b9550396bf6d278330281e8b709c/charset_normalizer-3.4.7-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:3dedcc22d73ec993f42055eff4fcfed9318d1eeb9a6606c55892a26964964e48", size = 226176, upload-time = "2026-04-02T09:27:00.437Z" },
+    { url = "https://files.pythonhosted.org/packages/6a/9f/130394f9bbe06f4f63e22641d32fc9b202b7e251c9aef4db044324dac493/charset_normalizer-3.4.7-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:64f02c6841d7d83f832cd97ccf8eb8a906d06eb95d5276069175c696b024b60a", size = 217723, upload-time = "2026-04-02T09:27:02.021Z" },
+    { url = "https://files.pythonhosted.org/packages/73/55/c469897448a06e49f8fa03f6caae97074fde823f432a98f979cc42b90e69/charset_normalizer-3.4.7-cp313-cp313-win32.whl", hash = "sha256:4042d5c8f957e15221d423ba781e85d553722fc4113f523f2feb7b188cc34c5e", size = 148085, upload-time = "2026-04-02T09:27:03.192Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/78/1b74c5bbb3f99b77a1715c91b3e0b5bdb6fe302d95ace4f5b1bec37b0167/charset_normalizer-3.4.7-cp313-cp313-win_amd64.whl", hash = "sha256:3946fa46a0cf3e4c8cb1cc52f56bb536310d34f25f01ca9b6c16afa767dab110", size = 158819, upload-time = "2026-04-02T09:27:04.454Z" },
+    { url = "https://files.pythonhosted.org/packages/68/86/46bd42279d323deb8687c4a5a811fd548cb7d1de10cf6535d099877a9a9f/charset_normalizer-3.4.7-cp313-cp313-win_arm64.whl", hash = "sha256:80d04837f55fc81da168b98de4f4b797ef007fc8a79ab71c6ec9bc4dd662b15b", size = 147915, upload-time = "2026-04-02T09:27:05.971Z" },
+    { url = "https://files.pythonhosted.org/packages/97/c8/c67cb8c70e19ef1960b97b22ed2a1567711de46c4ddf19799923adc836c2/charset_normalizer-3.4.7-cp314-cp314-macosx_10_15_universal2.whl", hash = "sha256:c36c333c39be2dbca264d7803333c896ab8fa7d4d6f0ab7edb7dfd7aea6e98c0", size = 309234, upload-time = "2026-04-02T09:27:07.194Z" },
+    { url = "https://files.pythonhosted.org/packages/99/85/c091fdee33f20de70d6c8b522743b6f831a2f1cd3ff86de4c6a827c48a76/charset_normalizer-3.4.7-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1c2aed2e5e41f24ea8ef1590b8e848a79b56f3a5564a65ceec43c9d692dc7d8a", size = 208042, upload-time = "2026-04-02T09:27:08.749Z" },
+    { url = "https://files.pythonhosted.org/packages/87/1c/ab2ce611b984d2fd5d86a5a8a19c1ae26acac6bad967da4967562c75114d/charset_normalizer-3.4.7-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:54523e136b8948060c0fa0bc7b1b50c32c186f2fceee897a495406bb6e311d2b", size = 228706, upload-time = "2026-04-02T09:27:09.951Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/29/2b1d2cb00bf085f59d29eb773ce58ec2d325430f8c216804a0a5cd83cbca/charset_normalizer-3.4.7-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:715479b9a2802ecac752a3b0efa2b0b60285cf962ee38414211abdfccc233b41", size = 224727, upload-time = "2026-04-02T09:27:11.175Z" },
+    { url = "https://files.pythonhosted.org/packages/47/5c/032c2d5a07fe4d4855fea851209cca2b6f03ebeb6d4e3afdb3358386a684/charset_normalizer-3.4.7-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bd6c2a1c7573c64738d716488d2cdd3c00e340e4835707d8fdb8dc1a66ef164e", size = 215882, upload-time = "2026-04-02T09:27:12.446Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/c2/356065d5a8b78ed04499cae5f339f091946a6a74f91e03476c33f0ab7100/charset_normalizer-3.4.7-cp314-cp314-manylinux_2_31_armv7l.whl", hash = "sha256:c45e9440fb78f8ddabcf714b68f936737a121355bf59f3907f4e17721b9d1aae", size = 200860, upload-time = "2026-04-02T09:27:13.721Z" },
+    { url = "https://files.pythonhosted.org/packages/0c/cd/a32a84217ced5039f53b29f460962abb2d4420def55afabe45b1c3c7483d/charset_normalizer-3.4.7-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3534e7dcbdcf757da6b85a0bbf5b6868786d5982dd959b065e65481644817a18", size = 211564, upload-time = "2026-04-02T09:27:15.272Z" },
+    { url = "https://files.pythonhosted.org/packages/44/86/58e6f13ce26cc3b8f4a36b94a0f22ae2f00a72534520f4ae6857c4b81f89/charset_normalizer-3.4.7-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:e8ac484bf18ce6975760921bb6148041faa8fef0547200386ea0b52b5d27bf7b", size = 211276, upload-time = "2026-04-02T09:27:16.834Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/fe/d17c32dc72e17e155e06883efa84514ca375f8a528ba2546bee73fc4df81/charset_normalizer-3.4.7-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:a5fe03b42827c13cdccd08e6c0247b6a6d4b5e3cdc53fd1749f5896adcdc2356", size = 201238, upload-time = "2026-04-02T09:27:18.229Z" },
+    { url = "https://files.pythonhosted.org/packages/6a/29/f33daa50b06525a237451cdb6c69da366c381a3dadcd833fa5676bc468b3/charset_normalizer-3.4.7-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:2d6eb928e13016cea4f1f21d1e10c1cebd5a421bc57ddf5b1142ae3f86824fab", size = 230189, upload-time = "2026-04-02T09:27:19.445Z" },
+    { url = "https://files.pythonhosted.org/packages/b6/6e/52c84015394a6a0bdcd435210a7e944c5f94ea1055f5cc5d56c5fe368e7b/charset_normalizer-3.4.7-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:e74327fb75de8986940def6e8dee4f127cc9752bee7355bb323cc5b2659b6d46", size = 211352, upload-time = "2026-04-02T09:27:20.79Z" },
+    { url = "https://files.pythonhosted.org/packages/8c/d7/4353be581b373033fb9198bf1da3cf8f09c1082561e8e922aa7b39bf9fe8/charset_normalizer-3.4.7-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:d6038d37043bced98a66e68d3aa2b6a35505dc01328cd65217cefe82f25def44", size = 227024, upload-time = "2026-04-02T09:27:22.063Z" },
+    { url = "https://files.pythonhosted.org/packages/30/45/99d18aa925bd1740098ccd3060e238e21115fffbfdcb8f3ece837d0ace6c/charset_normalizer-3.4.7-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:7579e913a5339fb8fa133f6bbcfd8e6749696206cf05acdbdca71a1b436d8e72", size = 217869, upload-time = "2026-04-02T09:27:23.486Z" },
+    { url = "https://files.pythonhosted.org/packages/5c/05/5ee478aa53f4bb7996482153d4bfe1b89e0f087f0ab6b294fcf92d595873/charset_normalizer-3.4.7-cp314-cp314-win32.whl", hash = "sha256:5b77459df20e08151cd6f8b9ef8ef1f961ef73d85c21a555c7eed5b79410ec10", size = 148541, upload-time = "2026-04-02T09:27:25.146Z" },
+    { url = "https://files.pythonhosted.org/packages/48/77/72dcb0921b2ce86420b2d79d454c7022bf5be40202a2a07906b9f2a35c97/charset_normalizer-3.4.7-cp314-cp314-win_amd64.whl", hash = "sha256:92a0a01ead5e668468e952e4238cccd7c537364eb7d851ab144ab6627dbbe12f", size = 159634, upload-time = "2026-04-02T09:27:26.642Z" },
+    { url = "https://files.pythonhosted.org/packages/c6/a3/c2369911cd72f02386e4e340770f6e158c7980267da16af8f668217abaa0/charset_normalizer-3.4.7-cp314-cp314-win_arm64.whl", hash = "sha256:67f6279d125ca0046a7fd386d01b311c6363844deac3e5b069b514ba3e63c246", size = 148384, upload-time = "2026-04-02T09:27:28.271Z" },
+    { url = "https://files.pythonhosted.org/packages/94/09/7e8a7f73d24dba1f0035fbbf014d2c36828fc1bf9c88f84093e57d315935/charset_normalizer-3.4.7-cp314-cp314t-macosx_10_15_universal2.whl", hash = "sha256:effc3f449787117233702311a1b7d8f59cba9ced946ba727bdc329ec69028e24", size = 330133, upload-time = "2026-04-02T09:27:29.474Z" },
+    { url = "https://files.pythonhosted.org/packages/8d/da/96975ddb11f8e977f706f45cddd8540fd8242f71ecdb5d18a80723dcf62c/charset_normalizer-3.4.7-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:fbccdc05410c9ee21bbf16a35f4c1d16123dcdeb8a1d38f33654fa21d0234f79", size = 216257, upload-time = "2026-04-02T09:27:30.793Z" },
+    { url = "https://files.pythonhosted.org/packages/e5/e8/1d63bf8ef2d388e95c64b2098f45f84758f6d102a087552da1485912637b/charset_normalizer-3.4.7-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:733784b6d6def852c814bce5f318d25da2ee65dd4839a0718641c696e09a2960", size = 234851, upload-time = "2026-04-02T09:27:32.44Z" },
+    { url = "https://files.pythonhosted.org/packages/9b/40/e5ff04233e70da2681fa43969ad6f66ca5611d7e669be0246c4c7aaf6dc8/charset_normalizer-3.4.7-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:a89c23ef8d2c6b27fd200a42aa4ac72786e7c60d40efdc76e6011260b6e949c4", size = 233393, upload-time = "2026-04-02T09:27:34.03Z" },
+    { url = "https://files.pythonhosted.org/packages/be/c1/06c6c49d5a5450f76899992f1ee40b41d076aee9279b49cf9974d2f313d5/charset_normalizer-3.4.7-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6c114670c45346afedc0d947faf3c7f701051d2518b943679c8ff88befe14f8e", size = 223251, upload-time = "2026-04-02T09:27:35.369Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/9f/f2ff16fb050946169e3e1f82134d107e5d4ae72647ec8a1b1446c148480f/charset_normalizer-3.4.7-cp314-cp314t-manylinux_2_31_armv7l.whl", hash = "sha256:a180c5e59792af262bf263b21a3c49353f25945d8d9f70628e73de370d55e1e1", size = 206609, upload-time = "2026-04-02T09:27:36.661Z" },
+    { url = "https://files.pythonhosted.org/packages/69/d5/a527c0cd8d64d2eab7459784fb4169a0ac76e5a6fc5237337982fd61347e/charset_normalizer-3.4.7-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:3c9a494bc5ec77d43cea229c4f6db1e4d8fe7e1bbffa8b6f0f0032430ff8ab44", size = 220014, upload-time = "2026-04-02T09:27:38.019Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/80/8a7b8104a3e203074dc9aa2c613d4b726c0e136bad1cc734594b02867972/charset_normalizer-3.4.7-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:8d828b6667a32a728a1ad1d93957cdf37489c57b97ae6c4de2860fa749b8fc1e", size = 218979, upload-time = "2026-04-02T09:27:39.37Z" },
+    { url = "https://files.pythonhosted.org/packages/02/9a/b759b503d507f375b2b5c153e4d2ee0a75aa215b7f2489cf314f4541f2c0/charset_normalizer-3.4.7-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:cf1493cd8607bec4d8a7b9b004e699fcf8f9103a9284cc94962cb73d20f9d4a3", size = 209238, upload-time = "2026-04-02T09:27:40.722Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/4e/0f3f5d47b86bdb79256e7290b26ac847a2832d9a4033f7eb2cd4bcf4bb5b/charset_normalizer-3.4.7-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:0c96c3b819b5c3e9e165495db84d41914d6894d55181d2d108cc1a69bfc9cce0", size = 236110, upload-time = "2026-04-02T09:27:42.33Z" },
+    { url = "https://files.pythonhosted.org/packages/96/23/bce28734eb3ed2c91dcf93abeb8a5cf393a7b2749725030bb630e554fdd8/charset_normalizer-3.4.7-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:752a45dc4a6934060b3b0dab47e04edc3326575f82be64bc4fc293914566503e", size = 219824, upload-time = "2026-04-02T09:27:43.924Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/6f/6e897c6984cc4d41af319b077f2f600fc8214eb2fe2d6bcb79141b882400/charset_normalizer-3.4.7-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:8778f0c7a52e56f75d12dae53ae320fae900a8b9b4164b981b9c5ce059cd1fcb", size = 233103, upload-time = "2026-04-02T09:27:45.348Z" },
+    { url = "https://files.pythonhosted.org/packages/76/22/ef7bd0fe480a0ae9b656189ec00744b60933f68b4f42a7bb06589f6f576a/charset_normalizer-3.4.7-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:ce3412fbe1e31eb81ea42f4169ed94861c56e643189e1e75f0041f3fe7020abe", size = 225194, upload-time = "2026-04-02T09:27:46.706Z" },
+    { url = "https://files.pythonhosted.org/packages/c5/a7/0e0ab3e0b5bc1219bd80a6a0d4d72ca74d9250cb2382b7c699c147e06017/charset_normalizer-3.4.7-cp314-cp314t-win32.whl", hash = "sha256:c03a41a8784091e67a39648f70c5f97b5b6a37f216896d44d2cdcb82615339a0", size = 159827, upload-time = "2026-04-02T09:27:48.053Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/1d/29d32e0fb40864b1f878c7f5a0b343ae676c6e2b271a2d55cc3a152391da/charset_normalizer-3.4.7-cp314-cp314t-win_amd64.whl", hash = "sha256:03853ed82eeebbce3c2abfdbc98c96dc205f32a79627688ac9a27370ea61a49c", size = 174168, upload-time = "2026-04-02T09:27:49.795Z" },
+    { url = "https://files.pythonhosted.org/packages/de/32/d92444ad05c7a6e41fb2036749777c163baf7a0301a040cb672d6b2b1ae9/charset_normalizer-3.4.7-cp314-cp314t-win_arm64.whl", hash = "sha256:c35abb8bfff0185efac5878da64c45dafd2b37fb0383add1be155a763c1f083d", size = 153018, upload-time = "2026-04-02T09:27:51.116Z" },
+    { url = "https://files.pythonhosted.org/packages/db/8f/61959034484a4a7c527811f4721e75d02d653a35afb0b6054474d8185d4c/charset_normalizer-3.4.7-py3-none-any.whl", hash = "sha256:3dce51d0f5e7951f8bb4900c257dad282f49190fdbebecd4ba99bcc41fef404d", size = 61958, upload-time = "2026-04-02T09:28:37.794Z" },
+]
+
 [[package]]
 name = "click"
 version = "8.3.2"
@@ -408,22 +481,21 @@ wheels = [
 
 [[package]]
 name = "huggingface-hub"
-version = "1.11.0"
+version = "0.36.2"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "filelock" },
     { name = "fsspec" },
-    { name = "hf-xet", marker = "platform_machine == 'AMD64' or platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64'" },
-    { name = "httpx" },
+    { name = "hf-xet", marker = "platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64'" },
     { name = "packaging" },
     { name = "pyyaml" },
+    { name = "requests" },
     { name = "tqdm" },
-    { name = "typer" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/dc/89/e7aa12d8a6b9259bed10671abb25ae6fa437c0f88a86ecbf59617bae7759/huggingface_hub-1.11.0.tar.gz", hash = "sha256:15fb3713c7f9cdff7b808a94fd91664f661ab142796bb48c9cd9493e8d166278", size = 761749, upload-time = "2026-04-16T13:07:39.73Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/7c/b7/8cb61d2eece5fb05a83271da168186721c450eb74e3c31f7ef3169fa475b/huggingface_hub-0.36.2.tar.gz", hash = "sha256:1934304d2fb224f8afa3b87007d58501acfda9215b334eed53072dd5e815ff7a", size = 649782, upload-time = "2026-02-06T09:24:13.098Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/37/02/4f3f8997d1ea7fe0146b343e5e14bd065fa87af790d07e5576d31b31cc18/huggingface_hub-1.11.0-py3-none-any.whl", hash = "sha256:42a6de0afbfeb5e022222d36398f029679db4eb4778801aafda32257ae9131ab", size = 645499, upload-time = "2026-04-16T13:07:37.716Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/af/48ac8483240de756d2438c380746e7130d1c6f75802ef22f3c6d49982787/huggingface_hub-0.36.2-py3-none-any.whl", hash = "sha256:48f0c8eac16145dfce371e9d2d7772854a4f591bcb56c9cf548accf531d54270", size = 566395, upload-time = "2026-02-06T09:24:11.133Z" },
 ]
 
 [[package]]
@@ -494,8 +566,8 @@ requires-dist = [
     { name = "python-magic", specifier = ">=0.4.27" },
     { name = "ruff", marker = "extra == 'dev'", specifier = ">=0.8" },
     { name = "sqlalchemy", extras = ["asyncio"], specifier = ">=2.0.36" },
-    { name = "surya-ocr", marker = "extra == 'ocr'", specifier = ">=0.9" },
-    { name = "torch", marker = "extra == 'ocr'", specifier = ">=2.4" },
+    { name = "surya-ocr", marker = "extra == 'ocr'", specifier = ">=0.17,<0.18" },
+    { name = "torch", marker = "extra == 'ocr'", specifier = ">=2.7" },
     { name = "uvicorn", extras = ["standard"], specifier = ">=0.32" },
 ]
 provides-extras = ["ocr", "dev"]
@@ -593,18 +665,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/68/a5/19d7aaa7e433713ffe881df33705925a196afb9532efc8475d26593921a6/mako-1.3.11-py3-none-any.whl", hash = "sha256:e372c6e333cf004aa736a15f425087ec977e1fcbd2966aae7f17c8dc1da27a77", size = 78503, upload-time = "2026-04-14T20:19:53.233Z" },
 ]
 
-[[package]]
-name = "markdown-it-py"
-version = "4.0.0"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "mdurl" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/5b/f5/4ec618ed16cc4f8fb3b701563655a69816155e79e24a17b651541804721d/markdown_it_py-4.0.0.tar.gz", hash = "sha256:cb0a2b4aa34f932c007117b194e945bd74e0ec24133ceb5bac59009cda1cb9f3", size = 73070, upload-time = "2025-08-11T12:57:52.854Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/94/54/e7d793b573f298e1c9013b8c4dade17d481164aa517d1d7148619c2cedbf/markdown_it_py-4.0.0-py3-none-any.whl", hash = "sha256:87327c59b172c5011896038353a81343b6754500a08cd7a4973bb48c6d578147", size = 87321, upload-time = "2025-08-11T12:57:51.923Z" },
-]
-
 [[package]]
 name = "markupsafe"
 version = "3.0.3"
@@ -668,15 +728,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/70/bc/6f1c2f612465f5fa89b95bead1f44dcb607670fd42891d8fdcd5d039f4f4/markupsafe-3.0.3-cp314-cp314t-win_arm64.whl", hash = "sha256:32001d6a8fc98c8cb5c947787c5d08b0a50663d139f1305bac5885d98d9b40fa", size = 14146, upload-time = "2025-09-27T18:37:28.327Z" },
 ]
 
-[[package]]
-name = "mdurl"
-version = "0.1.2"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/d6/54/cfe61301667036ec958cb99bd3efefba235e65cdeb9c84d24a8293ba1d90/mdurl-0.1.2.tar.gz", hash = "sha256:bb413d29f5eea38f31dd4754dd7377d4465116fb207585f97bf925588687c1ba", size = 8729, upload-time = "2022-08-14T12:40:10.846Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/b3/38/89ba8ad64ae25be8de66a6d463314cf1eb366222074cfda9ee839c56a4b4/mdurl-0.1.2-py3-none-any.whl", hash = "sha256:84008a41e51615a49fc9966191ff91509e3c40b939176e643fd50a5c2196b8f8", size = 9979, upload-time = "2022-08-14T12:40:09.779Z" },
-]
-
 [[package]]
 name = "mpmath"
 version = "1.3.0"
@@ -1434,16 +1485,18 @@ wheels = [
 ]
 
 [[package]]
-name = "rich"
-version = "15.0.0"
+name = "requests"
+version = "2.33.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "markdown-it-py" },
-    { name = "pygments" },
+    { name = "certifi" },
+    { name = "charset-normalizer" },
+    { name = "idna" },
+    { name = "urllib3" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/c0/8f/0722ca900cc807c13a6a0c696dacf35430f72e0ec571c4275d2371fca3e9/rich-15.0.0.tar.gz", hash = "sha256:edd07a4824c6b40189fb7ac9bc4c52536e9780fbbfbddf6f1e2502c31b068c36", size = 230680, upload-time = "2026-04-12T08:24:00.75Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/5f/a4/98b9c7c6428a668bf7e42ebb7c79d576a1c3c1e3ae2d47e674b468388871/requests-2.33.1.tar.gz", hash = "sha256:18817f8c57c6263968bc123d237e3b8b08ac046f5456bd1e307ee8f4250d3517", size = 134120, upload-time = "2026-03-30T16:09:15.531Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/82/3b/64d4899d73f91ba49a8c18a8ff3f0ea8f1c1d75481760df8c68ef5235bf5/rich-15.0.0-py3-none-any.whl", hash = "sha256:33bd4ef74232fb73fe9279a257718407f169c09b78a87ad3d296f548e27de0bb", size = 310654, upload-time = "2026-04-12T08:24:02.83Z" },
+    { url = "https://files.pythonhosted.org/packages/d7/8e/7540e8a2036f79a125c1d2ebadf69ed7901608859186c856fa0388ef4197/requests-2.33.1-py3-none-any.whl", hash = "sha256:4e6d1ef462f3626a1f0a0a9c42dd93c63bad33f9f1c1937509b8c5c8718ab56a", size = 64947, upload-time = "2026-03-30T16:09:13.83Z" },
 ]
 
 [[package]]
@@ -1502,15 +1555,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/e1/e3/c164c88b2e5ce7b24d667b9bd83589cf4f3520d97cad01534cd3c4f55fdb/setuptools-81.0.0-py3-none-any.whl", hash = "sha256:fdd925d5c5d9f62e4b74b30d6dd7828ce236fd6ed998a08d81de62ce5a6310d6", size = 1062021, upload-time = "2026-02-06T21:10:37.175Z" },
 ]
 
-[[package]]
-name = "shellingham"
-version = "1.5.4"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/58/15/8b3609fd3830ef7b27b655beb4b4e9c62313a4e8da8c676e142cc210d58e/shellingham-1.5.4.tar.gz", hash = "sha256:8dbca0739d487e5bd35ab3ca4b36e11c4078f3a234bfce294b0a0291363404de", size = 10310, upload-time = "2023-10-24T04:13:40.426Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/e0/f9/0595336914c5619e5f28a1fb793285925a8cd4b432c9da0a987836c7f822/shellingham-1.5.4-py2.py3-none-any.whl", hash = "sha256:7ecfff8f2fd72616f7481040475a65b2bf8af90a56c89140852d1120324e8686", size = 9755, upload-time = "2023-10-24T04:13:38.866Z" },
-]
-
 [[package]]
 name = "six"
 version = "1.17.0"
@@ -1703,22 +1747,23 @@ wheels = [
 
 [[package]]
 name = "transformers"
-version = "5.5.4"
+version = "4.57.6"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
+    { name = "filelock" },
     { name = "huggingface-hub" },
     { name = "numpy" },
     { name = "packaging" },
     { name = "pyyaml" },
     { name = "regex" },
+    { name = "requests" },
     { name = "safetensors" },
     { name = "tokenizers" },
     { name = "tqdm" },
-    { name = "typer" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/a5/1e/1e244ab2ab50a863e6b52cc55761910567fa532b69a6740f6e99c5fdbd98/transformers-5.5.4.tar.gz", hash = "sha256:2e67cadba81fc7608cc07c4dd54f524820bc3d95b1cabd0ef3db7733c4f8b82e", size = 8227649, upload-time = "2026-04-13T16:55:55.181Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/c4/35/67252acc1b929dc88b6602e8c4a982e64f31e733b804c14bc24b47da35e6/transformers-4.57.6.tar.gz", hash = "sha256:55e44126ece9dc0a291521b7e5492b572e6ef2766338a610b9ab5afbb70689d3", size = 10134912, upload-time = "2026-01-16T10:38:39.284Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/29/fb/162a66789c65e5afa3b051309240c26bf37fbc8fea285b4546ae747995a2/transformers-5.5.4-py3-none-any.whl", hash = "sha256:0bd6281b82966fe5a7a16f553ea517a9db1dee6284d7cb224dfd88fc0dd1c167", size = 10236696, upload-time = "2026-04-13T16:55:51.497Z" },
+    { url = "https://files.pythonhosted.org/packages/03/b8/e484ef633af3887baeeb4b6ad12743363af7cce68ae51e938e00aaa0529d/transformers-4.57.6-py3-none-any.whl", hash = "sha256:4c9e9de11333ddfe5114bc872c9f370509198acf0b87a832a0ab9458e2bd0550", size = 11993498, upload-time = "2026-01-16T10:38:31.289Z" },
 ]
 
 [[package]]
@@ -1738,21 +1783,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/f6/56/6113c23ff46c00aae423333eb58b3e60bdfe9179d542781955a5e1514cb3/triton-3.6.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:46bd1c1af4b6704e554cad2eeb3b0a6513a980d470ccfa63189737340c7746a7", size = 188397994, upload-time = "2026-01-20T16:01:14.236Z" },
 ]
 
-[[package]]
-name = "typer"
-version = "0.24.1"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "annotated-doc" },
-    { name = "click" },
-    { name = "rich" },
-    { name = "shellingham" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/f5/24/cb09efec5cc954f7f9b930bf8279447d24618bb6758d4f6adf2574c41780/typer-0.24.1.tar.gz", hash = "sha256:e39b4732d65fbdcde189ae76cf7cd48aeae72919dea1fdfc16593be016256b45", size = 118613, upload-time = "2026-02-21T16:54:40.609Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/4a/91/48db081e7a63bb37284f9fbcefda7c44c277b18b0e13fbc36ea2335b71e6/typer-0.24.1-py3-none-any.whl", hash = "sha256:112c1f0ce578bfb4cab9ffdabc68f031416ebcc216536611ba21f04e9aa84c9e", size = 56085, upload-time = "2026-02-21T16:54:41.616Z" },
-]
-
 [[package]]
 name = "typing-extensions"
 version = "4.15.0"
@@ -1774,6 +1804,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/dc/9b/47798a6c91d8bdb567fe2698fe81e0c6b7cb7ef4d13da4114b41d239f65d/typing_inspection-0.4.2-py3-none-any.whl", hash = "sha256:4ed1cacbdc298c220f1bd249ed5287caa16f34d44ef4e9c3d0cbad5b521545e7", size = 14611, upload-time = "2025-10-01T02:14:40.154Z" },
 ]
 
+[[package]]
+name = "urllib3"
+version = "2.6.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/c7/24/5f1b3bdffd70275f6661c76461e25f024d5a38a46f04aaca912426a2b1d3/urllib3-2.6.3.tar.gz", hash = "sha256:1b62b6884944a57dbe321509ab94fd4d3b307075e0c2eae991ac71ee15ad38ed", size = 435556, upload-time = "2026-01-07T16:24:43.925Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/39/08/aaaad47bc4e9dc8c725e68f9d04865dbcb2052843ff09c97b08904852d84/urllib3-2.6.3-py3-none-any.whl", hash = "sha256:bf272323e553dfb2e87d9bfd225ca7b0f467b919d7bbd355436d3fd37cb0acd4", size = 131584, upload-time = "2026-01-07T16:24:42.685Z" },
+]
+
 [[package]]
 name = "uvicorn"
 version = "0.44.0"