Release v0.9.3 · gordonmurray/firnflow

Summary

Query-time filtering: an optional filter on POST /ns/{namespace}/query (#84, contributed by @hev). A DataFusion SQL predicate applied through LanceDB as a prefilter, before vector ranking, full-text scoring, or hybrid fusion. A filtered vector query returns up to k neighbours that satisfy the predicate rather than post-filtering an already-selected top-k, so "filter": "id > 1000" or a predicate over _ingested_at narrows the search itself. The predicate uses the same SQL dialect as /list. The embedded Python package mirrors it with search(filter=...).
Error handling splits by cause. A bad predicate returns 400 (a parse error, an unknown column, a type mismatch, an unknown function, or an unsupported operator), while a storage or IO failure on a filtered query stays 500, so an object-store timeout is not misreported as a bad predicate. Some parseable-but-unimplemented predicate syntax makes Lance's planner panic; that is caught on the filtered path and returned as a 400 rather than unwinding the request.
Caching note. filter participates in the exact result cache key, and semantic_cache.enabled rejects filtered requests in v1. Predicates are cached by their exact text, so a volatile function such as now() or random() in a filter can keep returning a cached earlier result until the next write (tracked in #89); use a fixed bound instead.

This is Part 1 of #84. The scalar-attribute columns and the /facet endpoint from that issue are not in this release; #84 stays open to track them.