chroma-core/chroma · error · NotImplementedError

Building a ChromaLangchainEmbeddingFunction from config is n

Error message

Building a ChromaLangchainEmbeddingFunction from config is not supported. Please recreate the langchain embedding function and pass it to create_langchain_embedding.

What it means

build_from_config for the langchain bridge always raises NotImplementedError by design: arbitrary langchain embedding objects cannot be serialized (get_config only stores the class name and a placeholder note). When Chroma tries to rehydrate the EF from persisted config — typically when reopening a collection without supplying the function — this error tells you to reconstruct it manually.

Source

Thrown at chromadb/utils/embedding_functions/chroma_langchain_embedding_function.py:141

        else:
            # Cast to Sequence[str] to satisfy the type checker
            embeddings = self.embed_documents(cast(Sequence[str], input))

        # Convert to numpy arrays
        return [np.array(embedding, dtype=np.float32) for embedding in embeddings]

    @staticmethod
    def name() -> str:
        return "langchain"

    @staticmethod
    def build_from_config(
        config: Dict[str, Any]
    ) -> "EmbeddingFunction[Union[Documents, Images]]":
        # This is a placeholder implementation since we can't easily serialize and deserialize
        # langchain embedding functions. Users will need to recreate the langchain embedding function
        # and pass it to create_langchain_embedding.
        raise NotImplementedError(
            "Building a ChromaLangchainEmbeddingFunction from config is not supported. "
            "Please recreate the langchain embedding function and pass it to create_langchain_embedding."
        )

    def get_config(self) -> Dict[str, Any]:
        return {
            "embedding_function_class": self._embedding_function_class,
            "note": "This is a placeholder config. You will need to recreate the langchain embedding function.",
        }

    def validate_config_update(
        self, old_config: Dict[str, Any], new_config: Dict[str, Any]
    ) -> None:
        raise NotImplementedError(
            "Updating a ChromaLangchainEmbeddingFunction config is not supported. "
            "Please recreate the langchain embedding function and pass it to create_langchain_embedding."
        )

View on GitHub (pinned to aecdd12c8a)

Solutions

Recreate the langchain embedding at startup and pass it explicitly when reopening: get_collection(name, embedding_function=create_langchain_embedding(OpenAIEmbeddings(...))).
Cache/construct the wrapped langchain object once at boot and reuse it for every get_collection call.
If you need config-only persistence, switch to a native chromadb embedding function that supports build_from_config.

Example fix

# before (second process run)
col = client.get_collection("docs")  # tries build_from_config -> NotImplementedError

# after (every run)
from langchain_openai import OpenAIEmbeddings
from chromadb.utils.embedding_functions import create_langchain_embedding
col = client.get_collection(
    "docs",
    embedding_function=create_langchain_embedding(OpenAIEmbeddings(model="text-embedding-3-large")),
)

Defensive patterns

Strategy: fallback

Validate before calling

def get_collection_with_langchain_ef(client, name: str):
    from langchain_openai import OpenAIEmbeddings
    return client.get_collection(
        name,
        embedding_function=create_langchain_embedding(OpenAIEmbeddings(model="text-embedding-3-large")),
    )  # always supply the EF: langchain configs cannot be rebuilt

Try / catch

try:
    col = client.get_collection("docs")  # no EF supplied
except NotImplementedError as e:
    if "not supported" in str(e):
        col = client.get_collection(
            "docs",
            embedding_function=create_langchain_embedding(build_my_langchain_ef()),
        )
    else:
        raise

Prevention

Construct the langchain embedding once per process and pass it on every get_collection call.
Never rely on persisted EF config for the langchain bridge — it stores only a class-name placeholder.
If you need fully serializable functions, prefer native chromadb embedding functions.

When it happens

Trigger: A collection was created with a ChromaLangchainEmbeddingFunction (its persisted name is 'langchain'); later, get_collection is called without embedding_function= (or the system calls build_from_config on the stored config), triggering the unconditional raise.

Common situations: Process restart: app creates collection in run 1, then re-opens it in run 2 without re-passing the EF; server-side rehydration of collections whose EF config name is 'langchain'; teammates assuming the wrapper round-trips like native chromadb functions.

Related errors

AI-assisted analysis of chroma-core/chroma@aecdd12c8a (2026-08-16). Data as JSON: /api/errors/cc16ac74490c77e1. Report an issue: GitHub.