Chroma: A Developer-Friendly Embedded Store

Quick to start, fewer knobs

Chroma is an open-source embedding database aimed at developer productivity. You create a collection, add records (ids, embeddings or documents, metadata) and query with embeddings or text, optionally with a where metadata filter. It can run in-process (in memory or persisted to a folder) or as a server. If you give it documents without embeddings it calls an embedding function for you, which may download or call a model; in the example below we pass our own vectors so nothing is downloaded. Embedded stores like this are excellent for prototypes, tests and small single-user apps; for large multi-user production workloads compare it with the other families on scale, filtering, replication and operations.

Same ideas, different APIs

Most engines expose collections, upserts, filtered queries and deletes; compare them on your workload, not on a feature list.

Chroma: add, query, filter, delete, run

I ran this Python in a virtual environment with qdrant-client 1.19.1 (local in-memory mode) and chromadb 1.5.9 (in-memory client). Using cosine distance, the nearest records to [1,0,0] are a, d, b with small distances. A where filter for tenant acme and year 2025 leaves a and c. After deleting d the count is 3.

import chromadb

client = chromadb.EphemeralClient()                       # in-memory; we pass our own vectors, so no model download
col = client.create_collection("docs", metadata={"hnsw:space": "cosine"})
col.add(
    ids=["a", "b", "c", "d"],
    embeddings=[[0.9, 0.1, 0.0], [0.8, 0.2, 0.1], [0.1, 0.9, 0.1], [0.9, 0.0, 0.2]],
    documents=["leave policy 2025", "old leave policy", "travel and hotels", "beta confidential memo"],
    metadatas=[{"tenant": "acme", "year": 2025}, {"tenant": "acme", "year": 2022},
               {"tenant": "acme", "year": 2025}, {"tenant": "beta", "year": 2025}],
)
r = col.query(query_embeddings=[[1, 0, 0]], n_results=3)
print("no filter :", r["ids"][0], [round(d, 3) for d in r["distances"][0]])
r = col.query(query_embeddings=[[1, 0, 0]], n_results=3, where={"$and": [{"tenant": "acme"}, {"year": 2025}]})
print("with where:", r["ids"][0], r["documents"][0])
col.delete(ids=["d"])
print("count after delete:", col.count())

Output:

no filter : ['a', 'd', 'b'] [0.006, 0.024, 0.037]
with where: ['a', 'c'] ['leave policy 2025', 'travel and hotels']
count after delete: 3

Pass your own embeddings in production

Controlling the embedding model yourself keeps vectors reproducible and avoids surprise model downloads.

Quick check: What does Chroma do if you add documents without embeddings?

• Stores them as images
• Refuses all inserts
• Calls a configured embedding function to create them
• Deletes the collection