fix(genai): schema in prompt (#40)

src/ix/genai/ollama_client.py

@@ -162,22 +162,26 @@ class OllamaClient:
         """Map provider-neutral kwargs to Ollama's /api/chat body.
 
         Schema strategy for Ollama 0.11.8: we pass ``format="json"`` (loose
-        JSON mode) rather than the full Pydantic schema. The llama.cpp
-        structured-output implementation in 0.11.8 segfaults on schemas
-        involving ``anyOf``, ``$ref``, or ``pattern`` — which Pydantic v2
-        emits for Optional / nested-model / Decimal fields.
+        JSON mode) and bake the Pydantic schema into a system message
+        ahead of the caller's own system prompt. Rationale:
 
-        In loose JSON mode Ollama guarantees only syntactically-valid
-        JSON; we enforce the schema on our side by catching the Pydantic
-        ``ValidationError`` at parse time and raising IX_002_001. The
-        system prompt (built upstream in GenAIStep) already tells the
-        model what JSON shape to emit, so loose mode is the right
-        abstraction layer here.
+        * 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,
@@ -214,6 +218,34 @@ class OllamaClient:
         return out
 
 
+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.
 

tests/unit/test_ollama_client.py

@@ -85,7 +85,12 @@ class TestInvokeHappyPath:
         assert body_json["format"] == "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"},
         ]
@@ -119,7 +124,10 @@ 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"}
         ]