ragleap-vectorstores v0.6.0

Upgrade note

• Score thresholds: RedisBackend, ChromaBackend and LanceDBBackend now return similarity_score on the 0 to 1 scale (see Fixed). If you filter or alert on those scores, re-check your thresholds.

Fixed

• ChromaBackend.search_dense() and LanceDBBackend.search_dense() now return similarity_score on the same 0 to 1 scale as the other backends: 1 - distance / 2 (clamped), so identical vectors score 1.0, orthogonal 0.5 and opposite 0.0. Before they returned 1 - distance, the raw cosine in [-1, 1] (0.0 for orthogonal, -1.0 for opposite), the same inconsistency as issues 577 and the Redis fix. Behavior change: returned scores change; stored data is not affected. Docstrings that described the old convention were updated.
• RedisBackend.search_dense() now returns similarity_score on the same 0 to 1 scale as the other backends: 1 - distance / 2 (clamped), so identical vectors score 1.0, orthogonal 0.5 and opposite 0.0. Before it returned 1 - distance, the raw cosine in [-1, 1] (0.0 for orthogonal, -1.0 for opposite; live-verified on Redis Stack 7.4.7), so a caller filtering on similarity_score behaved differently per backend (same kind of inconsistency as issue 577). Behavior change: returned scores change; stored data is not affected.
• UpstashBackend.insert_chunk(): caller metadata can no longer overwrite the stored document_id, document_name, chunk_index, token_count or text (a key with one of those names used to replace the real value). Behavior change only for metadata that used those key names; other keys are stored as before.
• UpstashBackend.list_documents() now counts all chunks of a document: it follows the range() cursor until the prefix scan ends. Before, a single call capped chunk_count at 1000. Costs one extra range() request per further 1000 chunks of a listed document.

Verified

• The Redis score fix was checked live on a Redis Stack 7.4.7 container and runs in CI against redis-stack; the Chroma and LanceDB fixes against real embedded instances; the two Upstash fixes with mock-based tests only (no real Upstash index); the OpenSearch and Upstash module changes in the score fix are comment edits.

PyPI: https://pypi.org/project/ragleap-vectorstores/0.6.0/