pgchroma: Chroma's API on your Postgres¶

Chroma is a popular open-source vector database: AI apps use it to store embeddings and find the most similar documents (RAG, agent memory, semantic search). pgchroma is a server that speaks Chroma's HTTP API but keeps every collection in Postgres with pgvector. Your existing Chroma client code keeps working: you change the host, not the code.

Move off Chroma Cloud Star on GitHub

Chroma Cloud is shutting down

Chroma Cloud stops accepting writes on November 22, 2026 and stops serving reads on December 22, 2026 (announcement). Here is how to move your collections to Postgres.

Terminal recording: pgchroma migrate copies three collections from a Chroma server into Postgres and verifies them, then answers a query from unchanged chromadb code

import chromadb

# before: client = chromadb.CloudClient(api_key="ck-...", tenant="...", database="prod")
client = chromadb.HttpClient(host="localhost", port=8000, database="prod")

docs = client.get_collection("docs")  # everything below is unchanged
docs.query(query_texts=["how do refunds work?"], n_results=5, where={"lang": "en"})

Why pgchroma¶

No rewrite. The official chromadb clients for Python and JavaScript, and everything built on them (LangChain, LlamaIndex), work unchanged.
Tested against the real thing. The test suite runs the official client against a real Chroma 1.5.9 server and against pgchroma with the same calls and compares the results: 75 scenarios plus the JavaScript client.
Your Postgres. Supabase, Neon, RDS, Cloud SQL, Azure, Railway, or a laptop. Backups, replicas, access control and SQL joins come with it.
More than single-node Chroma. Chroma Cloud's Search API, sparse vectors (BM25, SPLADE) and hybrid search work on pgchroma. Single-node Chroma rejects them.
Stateless. Run one instance or twenty behind a load balancer.

Quickstart¶

docker run -p 8000:8000 -e DATABASE_URL="postgresql://user:pass@host:5432/db" ghcr.io/thejustas228/pgchroma

or, without Docker:

uvx --from git+https://github.com/TheJustas228/pgchroma pgchroma serve --database-url "$DATABASE_URL"

Then point any Chroma client at localhost:8000.

Performance¶

pgchroma vs Chroma 1.5.9 on 50,000 384-dimensional vectors: ingest 12,497 vs 1,274 records/s; filtered query 11.4 vs 37.2 ms; get with where 3.9 vs 11.3 ms; query 7.6 vs 5.3 ms

Measured with bench/bench.py: same client and data against both servers, one machine. Plain vector search is a little slower than Chroma; ingest and filtered search are much faster.

How it works¶

Your app's chromadb client talks the Chroma HTTP API to pgchroma serve, which stores everything in Postgres + pgvector. pgchroma migrate reads from Chroma Cloud and writes to the same Postgres.

Every collection is a Postgres table with a vector column, a jsonb metadata column and its own HNSW index. Filters compile to SQL. Nothing is stored outside Postgres.

pgchroma is MIT licensed and not affiliated with Chroma. If it saves you a migration, a star on GitHub helps other Chroma users find it.