Gate prediction-error flags on recurrence and fit; fix contextual_mismatch#47
Merged
Merged
Conversation
…match
Triaging the 46 live flags on LXC 122 showed ~0 actionable note defects.
The subsystem was measuring query difficulty, not note health:
- low_overlap fired on any single query whose top hit had low cosine —
bare-identifier lookups ("azvmsqlp02") that keyword-matched correctly,
abstract/meta sweeps, and cross-domain least-bad hits. 35 of 36 flags
came from one ad-hoc query each.
- contextual_mismatch flagged *correct* retrievals: it fired whenever the
#1 note was absent from the recall-limited in_context_notes boost set,
including exact-title hits with strong cosine (m365-copilot-mcp at sim
0.63, reddit-engagement-daemon at 0.69). Precision ~0%.
These false flags also demoted the correct notes in later searches via the
prediction-error demotion stage.
Changes:
- contextual_mismatch now requires the top note to also be a weak fit
(sim < CONTEXTUAL_MISMATCH_MAX_SIM = 0.45). A strong hit outside the
boost set is not a mismatch.
- Surfacing (vault_prediction_errors, CLI) and the retrieval demotion now
require >= PREDICTION_ERROR_MIN_OCCURRENCES (2) distinct events. Single
flags still accumulate toward the threshold but neither surface nor demote.
- New tests/test_prediction_errors.py exercises the detection branch
end-to-end (real in-memory sqlite): low_overlap fires below threshold,
no flag above, FTS-only hits skip detection, only deduped[0] is checked,
contextual_mismatch fires in-band and is suppressed for strong hits, and
the occurrence gate surfaces only recurrent notes.
On the live DB this collapses 46 surfaced flags to 1 (third-parties.md,
which genuinely surprised two distinct CSP/AOBO queries).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Triaging all 46 live prediction-error flags on LXC 122 found ~0 genuinely actionable note defects. The subsystem was measuring query difficulty, not note health.
low_overlapfired on any single query whosesearch: add--jsonflag for machine-readable output #1 result had low cosine. 35 of 36 flags came from one ad-hoc query each: bare-identifier lookups ("azvmsqlp02") that the keyword path matched correctly, abstract/meta sweeps ("open decisions verification needed TODO"), and cross-domain least-bad hits ("spline 3D shader"→ neurostack.md).contextual_mismatchflagged correct retrievals — it fired whenever thesearch: add--jsonflag for machine-readable output #1 note was absent from the recall-limitedin_context_notesboost set, including exact-title hits with strong cosine (m365-copilot-mcpat sim 0.63,reddit-engagement-daemonat 0.69). Root cause: the caller context label ("nyk-azure") isn't even a substring of the folder (nyk-europe-azure), so the set leans on a brittle tag/folder-cosine heuristic that excludes correctly-domiciled notes. Precision ≈ 0%.These false flags weren't inert — the prediction-error demotion stage down-weighted the (correct) flagged notes in later retrieval.
Changes
contextual_mismatchnow also requires the top note to be a weak fit (sim < CONTEXTUAL_MISMATCH_MAX_SIM = 0.45). A strong hit outside the boost set is not a mismatch.vault_prediction_errorsMCP + CLI) and the retrieval demotion now require ≥PREDICTION_ERROR_MIN_OCCURRENCES(2) distinct events. Single flags still accumulate toward the threshold but neither surface nor demote.tests/test_prediction_errors.pyexercises the detection branch end-to-end with a real in-memory sqlite DB (not MagicMock): fires below threshold, no flag above, FTS-only hits skip detection, onlydeduped[0]is checked,contextual_mismatchfires in-band and is suppressed for strong hits, and the occurrence gate surfaces only recurrent notes.Impact
On the live DB this collapses 46 surfaced flags → 1 (
third-parties.md, which genuinely surprised two distinct CSP/AOBO queries). Full suite: 575 passed, ruff clean.