Hybrid search on Postgres

Chroma Cloud added sparse vectors and the Search API for keyword and hybrid search. Single-node Chroma rejects both ("Sparse vector is not supported for local chroma"). pgchroma runs them on Postgres, scored the way Chroma's source code scores them.

BM25 with Chroma's own embedding function

# pip install chromadb snowballstemmer
import chromadb
from chromadb import K, Knn, Rrf, Schema, Search, SparseVectorIndexConfig
from chromadb.utils.embedding_functions import ChromaBm25EmbeddingFunction

client = chromadb.HttpClient(host="localhost", port=8000)  # pgchroma
schema = Schema().create_index(
    config=SparseVectorIndexConfig(
        source_key=K.DOCUMENT, embedding_function=ChromaBm25EmbeddingFunction(), bm25=True
    ),
    key="bm25",
)
tickets = client.create_collection("tickets", schema=schema)
tickets.add(ids=["1", "2"], documents=["my refund never arrived", "the tracking link is broken"])

# keyword search
tickets.search(Search().rank(Knn(query="refund", key="bm25", limit=10)))

# hybrid: dense + BM25 fused with reciprocal rank fusion
tickets.search(
    Search()
    .rank(
        Rrf(
            [
                Knn(query="where is my money?", return_rank=True, limit=100, default=1000),
                Knn(
                    query="where is my money?",
                    key="bm25",
                    return_rank=True,
                    limit=100,
                    default=1000,
                ),
            ]
        )
    )
    .limit(10)
    .select(K.DOCUMENT, K.SCORE)
)

The client computes the sparse vectors (on add from the document, on search from the query text); pgchroma stores them in the record's metadata and ranks them.

How scoring works

Matching Chroma's implementation:

• A sparse Knn scores a record as 1 - query · document (lower is better). Only records that share at least one dimension with the query are candidates.
• With bm25=True in the collection schema, each query weight is multiplied by ln(1 + (N - n + 0.5) / (n + 0.5)), where N is the number of records in the collection and n the number whose vector contains that dimension.
• Rrf is computed from each Knn's rank: -Σ weight / (k + rank). Records missing from a Knn get its default, or are dropped when default is not set.
• group_by with MinK / MaxK keeps the top k records per group, then re-sorts by score.

Performance

Candidates come from Postgres's jsonb (GIN) index on the metadata; every candidate is then scored exactly. On 12,000 support tickets with BM25 vectors a keyword search takes about 60 ms. There is no inverted index for sparse vectors yet, so very large sparse collections are slower than on Chroma Cloud.

SPLADE

Collections that used Chroma Cloud's hosted chroma-cloud-splade can switch to the same open model running locally; see hosted embedding functions.