Add ModelConfig.extra_body for raw provider pass-through

CHANGELOG.md

@@ -8,6 +8,7 @@
 - **S3 / object-storage monitoring.** `monitor.s3: list[S3MonitorEntry]` adds a polling watcher per bucket prefix alongside the existing local-directory watcher. Each entry has its own `poll_interval`, `include_patterns`, `ignore_patterns`, `delete_orphans`, and `storage_options`. The same `serve --monitor` flag enables both. Orphan deletion is per-entry (scoped via `uri LIKE 's3://bucket/prefix/%'`); other buckets and prefixes are never touched.
 - **`[s3]` optional extra** (`obstore>=0.9`). Required for `s3://` sources and the S3 watcher. Uses obstore — the Python binding to the same Rust `object_store` crate that LanceDB uses internally — so `monitor.s3[*].storage_options` accepts the same dict shape as `lancedb.storage_options`. Empty/missing options fall back to the AWS default credential chain.
 - **`scripts/run-integration-tests.sh`** — wraps `docker compose up --wait`, `pytest -m integration`, and tear-down so the SeaweedFS-backed integration suite is a one-liner.
+- **`ModelConfig.extra_body`**. Optional dict forwarded verbatim to `ModelSettings.extra_body`, the raw pass-through pydantic-ai exposes for openai/ollama/anthropic/groq. Lets configs reach provider-specific keys without haiku.rag modelling them — e.g. `extra_body: {chat_template_kwargs: {enable_thinking: false}}` to disable Qwen3 thinking on a vLLM endpoint, where the high-level `enable_thinking` flag is a no-op.
 
 ### Changed
 

docs/configuration/providers.md

@@ -29,6 +29,7 @@ qa:
 - **max_tokens**: Maximum tokens in response. Default: unset (provider default), except title generation (100).
 - **enable_thinking**: Control reasoning behavior (see below)
 - **base_url**: Custom endpoint for OpenAI-compatible servers (vLLM, LM Studio, etc.)
+- **extra_body**: Raw dict forwarded to the model SDK (see [Raw Provider Pass-through](#raw-provider-pass-through))
 
 ### Thinking Control
 
@@ -66,6 +67,27 @@ See the [Pydantic AI thinking documentation](https://ai.pydantic.dev/thinking/)
 - Enable for QA, research, complex reasoning, and mathematical problems
 - Disable for speed-critical applications, title generation, and simple tasks
 
+### Raw Provider Pass-through
+
+The `extra_body` setting takes a dict that haiku.rag forwards verbatim to the underlying model SDK as `ModelSettings.extra_body`. Use it to reach provider-specific keys that haiku.rag does not model with a dedicated field.
+
+**Example — disable Qwen3 thinking on vLLM:**
+
+```yaml
+qa:
+  model:
+    provider: openai
+    name: qwen3.6-35b
+    base_url: http://localhost:11430/v1
+    extra_body:
+      chat_template_kwargs:
+        enable_thinking: false
+```
+
+vLLM serves Qwen3 chat templates that read their thinking switch from `chat_template_kwargs.enable_thinking`. The high-level `enable_thinking` setting on the openai provider maps to vLLM's `reasoning_effort` parameter, which Qwen3 templates ignore, so the field is a no-op for this combination. `extra_body` reaches the chat template directly and disables thinking. With it off, Qwen3 returns the answer in `content` immediately instead of emitting a hidden reasoning trace first.
+
+**Provider support:** honored by openai, ollama, anthropic, and groq via pydantic-ai's `ModelSettings.extra_body`. Silently ignored by gemini and bedrock.
+
 ## Embedding Providers
 
 Embedding models require three settings: `provider`, `name`, and `vector_dim`. Optionally, use `base_url` for OpenAI-compatible servers.

haiku_rag_slim/haiku/rag/config/models.py

@@ -17,6 +17,11 @@ class ModelConfig(BaseModel):
         temperature: Sampling temperature (0.0 to 1.0+)
         max_tokens: Maximum tokens to generate
         vision: True if the model can interpret images. Default False.
+        extra_body: Raw dict forwarded verbatim to the model SDK as
+            `ModelSettings.extra_body`. Provider-side escape hatch for
+            keys haiku.rag doesn't model explicitly (e.g. vLLM's
+            `chat_template_kwargs.enable_thinking: false` for Qwen3).
+            Honored by openai/ollama/anthropic/groq; ignored by gemini/bedrock.
     """
 
     provider: str = "ollama"
@@ -27,6 +32,7 @@ class ModelConfig(BaseModel):
     temperature: float | None = None
     max_tokens: int | None = None
     vision: bool = False
+    extra_body: dict | None = None
 
 
 class EmbeddingModelConfig(BaseModel):

haiku_rag_slim/haiku/rag/utils.py

@@ -101,7 +101,11 @@ def apply_common_settings(
     Returns:
         Updated settings instance or None if no settings to apply
     """
-    if model_config.temperature is None and model_config.max_tokens is None:
+    if (
+        model_config.temperature is None
+        and model_config.max_tokens is None
+        and model_config.extra_body is None
+    ):
         return settings
 
     if settings is None:
@@ -115,6 +119,9 @@ def apply_common_settings(
     if model_config.max_tokens is not None:
         settings_dict["max_tokens"] = model_config.max_tokens
 
+    if model_config.extra_body is not None:
+        settings_dict["extra_body"] = model_config.extra_body
+
     return settings_dict
 
 

tests/test_utils.py

@@ -208,6 +208,47 @@ def test_get_model_openai_non_reasoning_model_ignores_thinking():
     assert result._settings is None
 
 
+def test_get_model_openai_extra_body_forwarded():
+    """`extra_body` on ModelConfig is forwarded to ModelSettings.extra_body.
+
+    pydantic-ai's OpenAI model branch reads `model_settings["extra_body"]`
+    and passes it verbatim to the OpenAI SDK. Enables vLLM-specific keys
+    like `chat_template_kwargs.enable_thinking` without coupling them to
+    the high-level `enable_thinking` flag.
+    """
+    extra = {"chat_template_kwargs": {"enable_thinking": False}}
+    model_config = ModelConfig(
+        provider="openai",
+        name="qwen3.6-35b",
+        base_url="http://localhost:11430/v1",
+        extra_body=extra,
+    )
+    result = get_model(model_config)
+    assert isinstance(result, OpenAIChatModel)
+    assert result._settings is not None
+    assert result._settings.get("extra_body") == extra
+
+
+def test_get_model_ollama_extra_body_forwarded():
+    """`extra_body` is forwarded through the Ollama (openai-compatible) branch too."""
+    extra = {"chat_template_kwargs": {"enable_thinking": False}}
+    model_config = ModelConfig(provider="ollama", name="qwen3", extra_body=extra)
+    result = get_model(model_config)
+    assert isinstance(result, OpenAIChatModel)
+    assert result._settings is not None
+    assert result._settings.get("extra_body") == extra
+
+
+def test_get_model_extra_body_absent_when_unset():
+    """No `extra_body` key appears on the settings when the config omits it."""
+    model_config = ModelConfig(provider="openai", name="gpt-4o-mini", temperature=0.3)
+    result = get_model(model_config)
+    assert isinstance(result, OpenAIChatModel)
+    # temperature triggers settings construction; extra_body should not be there.
+    assert result._settings is not None
+    assert "extra_body" not in result._settings
+
+
 @pytest.mark.skipif(not HAS_ANTHROPIC, reason="Anthropic not installed")
 def test_get_model_anthropic():
     """Test get_model returns AnthropicModel for Anthropic."""