feat(llm): add EMBEDDING_MODEL_CONFIG__INPUT_TYPE for asymmetric embedders

.env.template

@@ -21,6 +21,11 @@ LOG_LEVEL=INFO
 # EMBEDDING_MODEL_CONFIG__TRANSPORT=openai
 # EMBEDDING_MODEL_CONFIG__MODEL=text-embedding-3-small
 # EMBEDDING_MODEL_CONFIG__OVERRIDES__BASE_URL=
+# Asymmetric-encoder providers (e.g. some NVIDIA NIM embedding models like
+# `nvidia/llama-nemotron-embed-1b-v2`) require an `input_type` parameter and
+# will fail without it. Set this to the value the provider expects — usually
+# `query` for indexing/search use. Leave unset for plain OpenAI / Gemini.
+# EMBEDDING_MODEL_CONFIG__INPUT_TYPE=query
 # EMBEDDING_MODEL_CONFIG__OVERRIDES__API_KEY_ENV=
 
 # LANGFUSE_HOST=

src/config.py

@@ -301,6 +301,13 @@ class ConfiguredEmbeddingModelSettings(BaseModel):
     transport: EmbeddingTransport = "openai"
     overrides: ModelOverrideSettings = Field(default_factory=ModelOverrideSettings)
     dimensions_mode: EmbeddingDimensionsMode = "auto"
+    # Optional `input_type` to forward to the embedding endpoint. Required by
+    # some asymmetric-encoder providers — notably NVIDIA NIM models such as
+    # `nvidia/llama-nemotron-embed-1b-v2` which fail with HTTP 400 when this
+    # field is missing. Common values: "query", "passage" (NIM), or
+    # provider-specific equivalents. Default `None` preserves the historical
+    # behaviour for plain OpenAI / Gemini, which don't accept this parameter.
+    input_type: str | None = None
 
     @model_validator(mode="before")
     @classmethod
@@ -337,6 +344,7 @@ class EmbeddingModelConfig(BaseModel):
     transport: EmbeddingTransport = "openai"
     api_key: str | None = None
     base_url: str | None = None
+    input_type: str | None = None
 
     @model_validator(mode="before")
     @classmethod
@@ -459,6 +467,7 @@ def resolve_embedding_model_config(
         transport=configured.transport,
         api_key=api_key,
         base_url=configured.overrides.base_url,
+        input_type=configured.input_type,
     )
 
 

src/embedding_client.py

@@ -152,6 +152,9 @@ def __init__(
         self.model: str = config.model
         self.vector_dimensions: int = vector_dimensions
         self.send_dimensions: bool = send_dimensions
+        # Optional `input_type` to forward to the embedding endpoint. Required
+        # by some asymmetric-encoder providers (notably NVIDIA NIM models).
+        self.input_type: str | None = config.input_type
 
         if self.transport == "gemini":
             if not config.api_key:
@@ -237,6 +240,8 @@ async def _call_openai() -> list[float]:
             openai_kwargs: dict[str, Any] = {"model": self.model, "input": [query]}
             if self.send_dimensions:
                 openai_kwargs["dimensions"] = self.vector_dimensions
+            if self.input_type is not None:
+                openai_kwargs["input_type"] = self.input_type
             response = await openai_client.embeddings.create(**openai_kwargs)
             return self._validate_embedding_dimensions(response.data[0].embedding)
 
@@ -290,6 +295,8 @@ async def _embed_batch(batch: list[str] = batch) -> list[list[float]]:
                     }
                     if self.send_dimensions:
                         openai_kwargs["dimensions"] = self.vector_dimensions
+                    if self.input_type is not None:
+                        openai_kwargs["input_type"] = self.input_type
                     response = await self.client.embeddings.create(**openai_kwargs)
                     batch_embeddings.extend(
                         [
@@ -455,6 +462,8 @@ async def _call_provider() -> dict[str, dict[int, list[float]]]:
                 }
                 if self.send_dimensions:
                     openai_kwargs["dimensions"] = self.vector_dimensions
+                if self.input_type is not None:
+                    openai_kwargs["input_type"] = self.input_type
                 response = await self.client.embeddings.create(**openai_kwargs)
                 for item, embedding_data in zip(batch, response.data, strict=True):
                     result[item.text_id][item.chunk_index] = (

tests/llm/test_embedding_client.py

@@ -328,6 +328,27 @@ def test_resolve_send_dimensions_always_overrides_ada_rejecting_allowlist(
     assert s.resolve_send_dimensions() is True
 
 
+def test_input_type_default_is_none(monkeypatch: pytest.MonkeyPatch) -> None:
+    """By default no input_type is sent — preserves OpenAI / Gemini behaviour."""
+    s = _build_embedding_settings({}, monkeypatch)
+    assert s.MODEL_CONFIG.input_type is None
+
+
+def test_input_type_propagates_through_resolved_config(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """When configured, input_type is preserved end-to-end through resolution."""
+    from src.config import resolve_embedding_model_config
+
+    s = _build_embedding_settings(
+        {"EMBEDDING_MODEL_CONFIG__INPUT_TYPE": "query"},
+        monkeypatch,
+    )
+    assert s.MODEL_CONFIG.input_type == "query"
+    resolved = resolve_embedding_model_config(s.MODEL_CONFIG)
+    assert resolved.input_type == "query"
+
+
 def test_resolve_send_dimensions_never_returns_false_regardless(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None: