A Chromia vector store for LangChain.js

LangChain.js lets you swap vector stores behind one interface: add documents, delete them, run a similarity search, or turn the store into a retriever. We wrote a Chromia vector store for @langchain/community so that a LangChain.js app can keep its embeddings on a Chromia blockchain. The work lives in our fork of langchain-ai/langchainjs. All credit for the framework goes to the LangChain maintainers and contributors.

What we added

The integration is about 400 lines in libs/langchain-community/src/vectorstores/chromia.ts, plus package wiring, two example scripts (search and delete) and a docs notebook. It extends LangChain's VectorStore base class and talks to Chromia through postchain-client, the standard Chromia client library.

You construct it with any LangChain embeddings model and a postchain client:

1import { Chromia } from "@langchain/community/vectorstores/chromia";
2
3const vectorStore = new Chromia(embeddings, {
4  client: postchainClient,
5  numDimensions: embeddings.dimensions,
6});
Copy

How it maps to Chromia

Writes and reads go to operations and queries defined on the Chromia side:

• addDocuments embeds the text and addVectors submits an add_messages operation in a transaction through client.sendTransaction.
• delete submits a delete_messages operation with the ids to remove.
• similaritySearchVectorWithScore calls the query_closest_objects query with the query vector and a query template (get_messages_with_distance), then turns the results back into LangChain Document objects with a distance score and their metadata.

Because it implements the base class, the usual helpers work as expected: similaritySearch, similaritySearchWithScore, filtered search, fromTexts, fromDocuments, and asRetriever, including MMR retrieval.

A follow-up commit tightened the response handling and how stored metadata is parsed back into document metadata.

Upstream status

We opened the integration upstream as langchain-ai/langchainjs PR #8041. That PR was closed without being merged, so the code is not part of the published @langchain/community package. Anyone who wants to use it today can take the files from our fork, which is otherwise a plain mirror of upstream.

Why it matters to us

Chromia stores data in a relational model on chain, which makes it a reasonable place to keep agent memory that other parties can verify. Wrapping it as a standard LangChain vector store means an existing retrieval pipeline can point at Chromia by changing the store, without rewriting the rest of the chain.