Chroma

Who it is for

Teams building a knowledge assistant or any retrieval system over their own documents, policies, contracts or tickets. It suits projects that want a simple, well-documented store without operating a heavyweight search cluster. If you already run Postgres everywhere, the pgvector extension may be the quieter operational choice, and very large corpora with strict latency demands can justify a dedicated managed vector service instead.

Install it

pip install chromadb
# start a server:
chroma run --path ./chroma-data

Before production

The database is the easy part. Answer quality is decided by chunking, embedding choice, retrieval strategy and evaluation, and each of those needs testing against real questions from your team. Plan for re-indexing, because changing your embedding model later means re-embedding everything. Chroma does not enforce document-level permissions on its own, so if different users may see different documents, that filtering has to be built into the application layer around it.

Where Blash AI comes in

We build the production retrieval layer on top: parsing with Unstructured or Marker from this library, a chunking and metadata scheme that fits your documents, permission filtering, citations in every answer, and an evaluation set that measures accuracy before anyone trusts it. The result is a system whose answers you can check rather than a demo that impresses once.