fix(search): guard dual-substrate lexicon levers (opt-in) + index .asc AsciiDoc#40
Merged
Conversation
added 3 commits
June 19, 2026 16:14
The lexicon levers (PR #38/#39) shipped on-by-default but regress retrieval on external repos with natural queries: neutral-to-negative on rustfmt/Go/flask/TS/prose (e.g. prose MRR 0.336->0.166, tanstack 0.104->0.007 on the same base). The wins were real but confined to engineered zero-overlap probes and a dense base. - query_index always computes the baseline ranking as a safety floor; expansion is DF-gated (drops common-word noise) and fused base-weighted (supplements but never evicts a confident base hit); rerank is gated to code-shaped queries. The levers reach parity instead of cratering. - levers are now opt-in (ADEN_LEXICON_ON); default is the baseline. - add scripts/lexicon_ab_bench.py (external A/B harness) and an external prose query set so re-enabling by default is gated on real repos.
filter.rs SECRET_EXTS listed "asc", so AsciiDoc books using the .asc extension (Pro Git and many doc repos) indexed empty with a green health score. .asc is dual-use (AsciiDoc + PGP armor); judge by content, not name: drop it from the skip-list and add a PGP-private-key content check so a real armored key in a .asc is still caught while prose indexes.
The blast-radius eval (graph caller-edges vs a text-scan ground truth of
NAME( call sites, file-level) existed only as a stale .pyc. Reconstruct it
as a committed, reusable harness that points at any external repo and
SEPARATES two concerns:
- understand name-resolution accuracy (did `understand NAME` land on the
exact symbol, or fuzzy-match a substring superset), and
- blast-radius precision/recall on the correctly-resolved symbols.
Auto-discovers curated callable gold symbols (no dunders/generic names,
called-but-not-ubiquitous), excludes test dirs to match aden's production
extraction scope. On flask: resolution 0.95, blast-radius P0.44/R0.61 —
the method-call-resolution gap depresses recall on OO/Python vs the 0.99
on free-function Rust code.
4 tasks
RioPlay
added a commit
that referenced
this pull request
Jun 20, 2026
commands.adoc: - asm --depth default: 3 → 2 (matches code) - mcp install: add missing --surface <essential|standard|full> flag - viz: add missing --scope and --resolution flags; remove bogus -j alias - heal: add structured flag table (--propose/--fix/--gc/--since/--apply/--watch) - status: document savings estimate output block ai-integration.adoc: - Replace two-tier Core/Extended model with three-tier Essential/Standard/Full - Fix tool assignments: search/list/communities/impact-diff were in wrong tier - Fix env var: ADEN_MCP_FULL=1 → ADEN_MCP_SURFACE=standard|full (legacy alias noted) - Document --surface flag at install time retrieval-levers.adoc: - Fix polarity: auto-gating is OFF by default, opt-in via ADEN_LEXICON_ON - Reframe ADEN_LEXICON_OFF as kill switch, not primary disable mechanism - Fix rerank trigger: code_anchor_fraction → query_looks_codey (query text only) - Fix NL-over-code behavior: expands only, does not rerank - Document PR #39/PR #40 revert history in status section architecture.adoc: - Add aden-paths node to Mermaid crate diagram - Remove duplicate aden-mcp Phase 2 row (already shipped as Phase 0) security-model.adoc: - Replace CanPerform (not a valid EdgeType) with Invokes in semantics example - Update malicious-contract-injection threat: moot since store-first (ADR-003) .agent/quick-ref.adoc: - Add --features watch caveat to aden watch entry crates/aden-cli/src/commands/init.rs: - Fix misleading "Knowledge graph built in .aden/store." message; store is in per-user data dir since ADR-003 docs/adr-008-current-implementation-state.adoc (new): - Document current scope extensions beyond original ADRs: Wave 3 edges (Supersedes/Justifies/AssociatedWith), GEN_LOGIC_VERSION=4, store-first architecture, MCP three-tier surface, dual-substrate retrieval (opt-in) Co-authored-by: RioPlay (Ernest Hamblen) <rioplay@rioplay.dev>
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.
Summary
The dual-substrate lexicon levers (PR #38/#39) shipped on by default, but external benchmarking showed they regress retrieval on every external repo with natural queries — e.g. prose MRR 0.336→0.166, tanstack 0.104→0.007 on the same base. The original wins were real but confined to engineered zero-overlap probes and a dense base, which the product path doesn't match. This PR makes the levers safe + opt-in, and fixes a related indexing bug found while building the prose benchmark.
fix(search)— guard the levers, flip to opt-in.query_indexnow always computes the baseline ranking as a safety floor; expansion is DF-gated (drops common-word noise) and fused base-weighted (can supplement but never evict a confident base hit); the PPMI rerank is gated to code-shaped queries. Levers are now opt-in viaADEN_LEXICON_ON— the default is the plain baseline (the best of everything measured). Guarded, enabling them reaches parity instead of cratering.fix(index)— index.ascAsciiDoc..ascwas inSECRET_EXTS(PGP-armor collision), so AsciiDoc books that use it (Pro Git, many docs repos) indexed empty with a green health score. It's dual-use, so it's now content-gated: a real armored PRIVATE key in a.ascis still caught bycontent_has_high_confidence_secret; prose indexes normally.test(bench)— addsscripts/lexicon_ab_bench.py(external ON/OFF A/B harness) + an external prose query set, so any future "on by default" decision is gated on real repos, not in-tree fixtures.Test plan
cargo test --workspacegreen (via the pre-push ci-check gate)cargo clippy --workspacecleancargo fmt --allappliedLicense checklist