Dev Tools

Spoiler: Chroma 1.0 dropped and the AI Twitter algorithm completely missed it. Which is wild, because the prototyping DB everyone learned RAG on just grew up.

The Setup

The pip install is still one line, the API is still the one you remember, and the migration from 0.5 to 1.0 is "change the import" levels of painless. But now it actually has hybrid search built in and persistence that doesn't corrupt itself when you ctrl-C.

import chromadb

client = chromadb.PersistentClient(path="./chroma_store")

collection = client.get_or_create_collection(
    name="claims",
    metadata={"hnsw:space": "cosine"},
)

collection.add(
    ids=["c1", "c2", "c3"],
    documents=[
        "hail damage to colorbond roofing in QLD",
        "storm water entered ceiling cavity",
        "wind-driven rain through window seals",
    ],
    metadatas=[{"region": "qld"}, {"region": "nsw"}, {"region": "qld"}],
)

The Money Pattern

The new hybrid query combines vector similarity with BM25 keyword scoring out of the box. You no longer have to bolt on a separate full-text index for entity-style queries. One call, both signals.

results = collection.query(
    query_texts=["roof leak after hailstorm in Brisbane"],
    n_results=5,
    where={"region": "qld"},
    # 1.0 ships native hybrid scoring
    include=["documents", "distances", "metadatas"],
)

for doc, dist in zip(results["documents"][0], results["distances"][0]):
    print(f"{dist:.3f}  {doc}")

The Catch

It's still primarily an in-memory DB with disk persistence bolted on the side. Past a few million vectors you'll feel the seams — long startup times, RAM ballooning, the works. And Chroma Cloud is in preview, so don't bet your prod stack on it yet.

The Verdict

For prototypes, internal tools, and anything with under a million vectors, Chroma 1.0 is genuinely great now. The hybrid search alone is enough reason to upgrade. Bigger workloads still want Qdrant or pgvector, but the "learning RAG on a Saturday" tier just got a serious glow-up.