feat: support Voyage embeddings via MongoDB Atlas (mongodb/ prefix)

EMBEDDINGS.md

@@ -135,6 +135,18 @@ envs:
   VOYAGE_API_KEY: your-api-key-here
 ```
 
+To access Voyage models through [MongoDB Atlas](https://www.mongodb.com/docs/voyageai/api-and-clients/),
+use the `mongodb/` prefix. Requests are routed to `https://ai.mongodb.com/v1` using your
+Atlas model API key (still read from `VOYAGE_API_KEY`):
+
+```yaml
+embedding:
+  provider: litellm
+  model: mongodb/voyage-4-large
+envs:
+  VOYAGE_API_KEY: your-atlas-model-api-key
+```
+
 For more information, see the [LiteLLM Providers Documentation](https://docs.litellm.ai/docs/providers).
 
 [Back to top](#table-of-contents)

src/cocoindex_code/embedder_defaults.py

@@ -80,7 +80,8 @@ class DefaultParamsEntry(NamedTuple):
     ),
     DefaultParamsEntry(
         "litellm",
-        re.compile(r"voyage/.+"),
+        # ``mongodb/`` routes Voyage models through MongoDB Atlas; same params.
+        re.compile(r"(voyage|mongodb)/.+"),
         {"input_type": "document"},
         {"input_type": "query"},
     ),

src/cocoindex_code/litellm_embedder.py

@@ -18,6 +18,23 @@
 _RATE_LIMIT_DELAY_RE = re.compile(r"Please try again in ([0-9.]+)(ms|s)", re.IGNORECASE)
 _MAX_RATE_LIMIT_RETRIES = 6
 
+# MongoDB Atlas serves Voyage AI embedding models through an endpoint that
+# speaks the native Voyage API, so a ``mongodb/<voyage-model>`` model name is
+# routed via LiteLLM's ``voyage/`` provider with ``api_base`` pointed at Atlas.
+_MONGODB_PREFIX = "mongodb/"
+_MONGODB_API_BASE = "https://ai.mongodb.com/v1"
+
+
+def _resolve_mongodb_model(model: str) -> tuple[str | None, str]:
+    """Map a ``mongodb/`` model name to a Voyage model plus the Atlas base URL.
+
+    Returns ``(api_base, model)``. For non-MongoDB models ``api_base`` is ``None``
+    and the model is returned unchanged.
+    """
+    if model.startswith(_MONGODB_PREFIX):
+        return _MONGODB_API_BASE, "voyage/" + model[len(_MONGODB_PREFIX) :]
+    return None, model
+
 
 def _get_rate_limit_delay(exc: Exception, attempt: int) -> float | None:
     message = str(exc)
@@ -39,7 +56,9 @@ class PacedLiteLLMEmbedder(LiteLLMEmbedder):
     """LiteLLM embedder that serializes requests and paces them when configured."""
 
     def __init__(self, model: str, *, min_interval_ms: int | None = None, **kwargs: Any) -> None:
+        api_base, model = _resolve_mongodb_model(model)
         super().__init__(model, **kwargs)
+        self._api_base = api_base
         self._min_request_interval_seconds = max(0.0, float(min_interval_ms or 0) / 1000.0)
         self._request_lock: asyncio.Lock | None = None
         self._next_request_at: float = 0.0
@@ -85,6 +104,10 @@ async def run_embedding_request(self, *, input: list[str], **kwargs: Any) -> Any
             if not self._model.startswith(("voyage/", "bedrock/")):
                 kwargs["encoding_format"] = "float"
                 kwargs["drop_params"] = True
+
+            if self._api_base is not None:
+                kwargs.setdefault("api_base", self._api_base)
+
             response = await self._aembedding_with_rate_limit_retries(
                 model=self._model,
                 input=input,

tests/test_embedder_defaults.py

@@ -35,6 +35,12 @@ def test_lookup_defaults_regex_match_voyage() -> None:
     assert query == {"input_type": "query"}
 
 
+def test_lookup_defaults_regex_match_mongodb_voyage() -> None:
+    indexing, query = lookup_defaults("litellm", "mongodb/voyage-4")
+    assert indexing == {"input_type": "document"}
+    assert query == {"input_type": "query"}
+
+
 def test_lookup_defaults_regex_match_cohere() -> None:
     indexing, query = lookup_defaults("litellm", "cohere/embed-english-v3.0")
     assert indexing == {"input_type": "search_document"}

tests/test_litellm_embedder.py

@@ -94,3 +94,27 @@ async def fake_aembedding(*, model: str, input: list[str], **kwargs: Any) -> Any
 
     assert "encoding_format" not in seen_kwargs
     assert "drop_params" not in seen_kwargs
+
+
+@pytest.mark.asyncio
+async def test_mongodb_model_routes_to_voyage_with_atlas_api_base(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    seen_model: list[str] = []
+    seen_kwargs: dict[str, Any] = {}
+
+    async def fake_aembedding(*, model: str, input: list[str], **kwargs: Any) -> Any:
+        seen_model.append(model)
+        seen_kwargs.update(kwargs)
+        return SimpleNamespace(data=[{"embedding": [1.0, 2.0]}])
+
+    monkeypatch.setattr("cocoindex_code.litellm_embedder.litellm.aembedding", fake_aembedding)
+
+    embedder = PacedLiteLLMEmbedder("mongodb/voyage-4")
+    await embedder.run_embedding_request(input=["hello"])
+
+    assert seen_model == ["voyage/voyage-4"]
+    assert seen_kwargs["api_base"] == "https://ai.mongodb.com/v1"
+    # Routed as a native Voyage provider, so encoding_format is not forced.
+    assert "encoding_format" not in seen_kwargs
+    assert "drop_params" not in seen_kwargs