Compatibility

Tested with the official clients chromadb 1.5.9 (Python) and chromadb 3.5.0 (npm), against a real Chroma 1.5.9 server, on Postgres 16 and 17 with pgvector 0.6, 0.7 and 0.8, in CI on every push (the core suite also runs on Postgres 15).

Tenants, databases, collections (create, get, get-or-create, list, count, modify, delete, fork)	✅
add, upsert, update, delete, get, query, count, peek	✅
where: $eq $ne $gt $gte $lt $lte $in $nin $and $or, list $contains / $not_contains	✅ same results as Chroma, including its int/float rules
where_document: $contains $not_contains $regex $not_regex $and $or	✅
include, limit / offset, querying by ids	✅
Distance spaces l2, cosine, ip (same distance values as Chroma)	✅
Collection configuration, embedding-function config, metadata, legacy hnsw:space metadata	✅
Base64 embedding transport, max batch size, pre-flight-checks	✅
Token auth (x-chroma-token or Authorization: Bearer)	✅
Search API: Search().where().rank(Knn / Rrf / arithmetic).limit().select(), #id / #document filters	✅
Sparse vectors (SparseVector metadata), sparse Knn, BM25 weighting (bm25=True), hybrid Rrf of dense + sparse	✅ scored like Chroma Cloud (1 - query·document, BM25 IDF over the collection), following Chroma's source; no self-hosted Chroma can run these to compare against
Search API: group_by with MinK / MaxK (top k per group)	✅ following Chroma's source
Attached functions, quotas, Chroma Cloud billing APIs	❌
Pre-1.0 clients (v1 API)	❌ answers 410 with an upgrade hint

Known differences

• delete(ids=[...]) returns the exact number of records deleted. Chroma returns len(ids) even for ids that do not exist.
• $regex uses Postgres regular expressions instead of Rust's regex crate. Common syntax (classes, anchors, alternation, (?i)) behaves the same; exotic features may not.
• Approximate search is pgvector's HNSW, not Chroma's HNSW or SPANN, so for large collections the last few of the top-k results can differ, the same way two runs of two different ANN indexes differ. Small collections are searched exactly.
• Chroma Cloud collections configured for SPANN are served with HNSW.
• A float metadata value of 1e16 or more is written by Python in exponent notation and comes back from Postgres as an integer.