Repository navigation
search_graph BM25: label filter applies after the FTS5 candidate window #2386
Description
The label filter on the BM25 path is applied after the 2000-row FTS5 candidate window (BM25_INNER_LIMIT, src/mcp/mcp.c ~3778; inner query ~4073), not inside candidate selection.
Effect: when a query token is common but the requested label is rare among its matches, the matching rows of that label can fall outside the window. The response then reports total_relation: gte / candidate_window_saturated, but can contain 0 rows, even though matching nodes exist.
Possible fix: filter on the label column of nodes_fts inside the FTS MATCH itself (or pre-select candidates per label), so the window is filled with rows of the requested label only.
Test: index a fixture where one token matches more than 2000 nodes of one label and a few of another; query with the rare label and assert those rows come back.
Activity
harshitaajoshi commentedon Sep 28, 2026
Hi, I've opened #2423 for this.
The label now goes into the FTS5 MATCH itself (
(<tokens>) AND label : "<Label>"), so the 2000 row candidate window only fills with rows of the requested label. The existingn.label = ?check stays as the exact filter, and the count query and saturation probe use the same match string, sototalandcandidate_window_saturateddescribe the filtered window.Two regression tests are included, one of which reproduces the issue (2001 Functions crowding out 3 Classes, which returns 0 rows on current main). Full suite and
lint-cipass locally. The PR description has details on a small ranking safeguard.Happy to adjust anything.
docs(fork): roadmap + community alignment sweep — phase-2 index econo…
fix(mcp): apply search_graph label filter inside the BM25 candidate w…
DeusData commentedon Oct 19, 2026
Thank you for opening #2423 and for the rare-label crowding regression. The count and saturation semantics are useful parts of the proposed change, alongside the returned rows. I have no new verified review or landing status to report in this issue.
fix(mcp): serve every named label in BM25 search_graph, never a silen…
DeusData commentedon Oct 25, 2026
Thank you for your patience and the filtered-window regression in #2423. Count/saturation behavior remains part of the maintainer review. Sorry for the delay; I’ll follow up soon.