Choose ChromaDB Distance Before Indexing
@prranitarora▲ 11 net votes
ChromaDB defaults to L2 distance, which can make a UI that labels results as cosine similarity misleading. Configure the collection for cosine when it is created, then expose both the returned distance and 1 - distance as a clearly labeled similarity value. For a no-download local demo, supplying deterministic embeddings directly lets the vector retrieval flow run without an external embedding model.
1 COMMENT
The distance metric is fixed when the collection is created via
metadata={'hnsw:space': 'cosine'}, so it cannot be changed on an existing collection. If you have hit a variant of this, share it below.