Cordis798 62de7f3425
perf(knowledge): bound concurrent WeKnora document checks (#7707)
Co-authored-by: Cordis798 <284918518+Cordis798@users.noreply.github.com>
2026-10-30 07:08:01 +08:00

2.1 KiB

WeKnora read-only extension

Keep provider-specific code in this independently packaged extension. Use registry.knowledge_provider and the public deerflow.knowledge contracts; core owns scope admission, formatting, budgets and citation evidence.

API reference: WeKnora v0.8.3 at 3e22822c1d2397be730c458c7539d0eb8c113a24. Preserve per-request operator bindings and upstream authorization. Never turn missing bindings, invalid selections, empty allowlists, or upstream failures into an unrestricted search. Use KB-bound hybrid-search with optional knowledge_ids and disabled context enrichment. The global search endpoint rejects document targets for KB-restricted keys and unions base/document target lists. Never broaden credentials to work around that. Verify result membership. Reject malformed document totals and detail shapes as provider errors; reserve permission errors for resource identity, membership and searchability failures. Dataset IDs derive from endpoint/tenant/native base, not credentials or labels.

Tests live in backend/tests/test_weknora_extension*.py; live tests require explicit opt-in and an already indexed instance. Do not fabricate a live result. See README for version references, configuration and the authorization model.

Optional entries_search uses KB-bound Wiki search and published pages only. Check capabilities.wiki and returned base membership; reject document filters before HTTP. Wiki responses use a separate pages envelope. Never advertise entity relations based on the Wiki page-link graph.

Wiki lexical search retries only successful empty pages with up to four distinct whitespace-separated keywords of at least two characters. Keep retries in the same authorized base, validate every response, deduplicate title/excerpt pairs, and never retry provider/auth/schema errors. Nonempty phrase results stay intact.

Batch document checks at four per search; retain dataset/ID request-local keys and hit order. Drain sibling tasks on errors/cancellation. Validate selected documents before search, and only the consumed hit prefix before returning it.