Context
attune-rag completed an alias-expansion-sweep (docs/specs/alias-expansion-sweep/) that closed a 41pp R@3 gap on paraphrased developer queries against the bundled help corpus. The mechanism:
- Diagnostic-1 (attune-rag/docs/specs/embedding-retriever/diagnostic-1.md) measured
KeywordRetriever collapsing on paraphrased queries (P@1 97.5% → 11.25%, R@3 100% → 28.75%).
- Diagnostic-3 (diagnostic-3.md) proved that hand-authored multi-token aliases on a single feature closed +50pp R@3 with zero baseline regression and zero new dependency.
- The sweep then applied that lever across 6 feature clusters (PRs #94–#99), landing paraphrased R@3 at 70% exactly (M13 acceptance criterion).
These aliases currently live in attune-rag/src/attune_rag/corpus/aliases_override.json as a temporary home — that override mechanism was built so the sweep could ship without requiring a new attune-help release per cluster. The long-term home is attune-help frontmatter so the aliases ship with the templates themselves.
What this issue tracks
Promote the 89 proven aliases (across 6 templates) from attune-rag's override file into upstream attune-help frontmatter. One attune-help release; six file edits; no behavior change. After the upstream release lands and attune-rag pins to the new attune-help version, attune-rag's aliases_override.json entries become redundant and can be deleted.
Inventory — 89 aliases across 6 templates
Source: attune-rag/src/attune_rag/corpus/aliases_override.json
concepts/tool-bug-predict.md (20 aliases)
Append to existing frontmatter aliases: list:
- dangerous code
- harmful patterns
- weak points
- fragile code
- code landmines
- danger zones
- what could go wrong
- silent failures
- fails silently
- hidden bugs
- PR risk review
- dangerous PR
- diff risk
- diff bite
- worrisome code
- shaky code
- commit risk
- production risk
- code that bites
- review risky changes
concepts/tool-security-audit.md (14 aliases)
Frontmatter currently has empty aliases: []. Add:
- unsafe input
- input handling
- leaked credentials
- leaked tokens
- exposed secrets
- attackers break in
- attack surface
- exploit surface
- exploit code
- security check
- service compromised
- compromised system
- risky function calls
- potentially attackable
concepts/tool-release-prep.md (14 aliases)
Append to existing frontmatter aliases: list:
- before pushing
- before pushing out
- before release
- push to users
- push to the world
- push code out
- push package
- push out the release
- ship to production
- ship to users
- release gate
- release checklist
- version bump
- tag a release
concepts/tool-smart-test.md (14 aliases)
Frontmatter currently has empty aliases: []. Add:
- safety nets
- build safety nets
- safety net for code
- untouched module
- untouched code
- untested code
- shore up coverage
- shore up tests
- functions need assertions
- assertions for functions
- missing tests
- test coverage gaps
- no tests yet
- code with no tests
concepts/tool-fix-test.md (13 aliases)
Append to existing frontmatter aliases: list:
- suite is red
- red test suite
- tests are red
- red after merge
- broken after merge
- figure out failures
- figure out why tests fail
- diagnose test failures
- diagnose failures
- blocking my merges
- blocking merges
- merges blocked
- fix broken suite
concepts/tool-code-quality.md (15 aliases)
Frontmatter currently has empty aliases: []. Add:
- once-over module
- give once-over
- module once-over
- quick once-over
- evaluate craftsmanship
- evaluate code changes
- code craftsmanship
- raise the bar
- raise code bar
- raise quality bar
- clean module
- keep code clean
- module clean code
- code health check
- polish code
Constraints to honor when promoting
- Global alias uniqueness. attune-help's frontmatter parser (or its consumers like attune-rag's
DirectoryCorpus) raises DuplicateAliasError if any alias appears on two templates. The 89 aliases above were chosen so they don't collide with each other or with existing frontmatter aliases as of attune-help 0.11.0. A pre-merge check should re-verify against whatever main is at promotion time.
- Each alias must be 2+ tokens after stemming. attune-rag's
KeywordRetriever.MIN_ALIAS_OVERLAP = 2 requires at least two query-token overlaps before crediting an alias. Single-token aliases get silently dropped. Every alias above was pre-validated with attune_rag.retrieval._tokenize.
- Stemmer-aware authoring. A few real footguns from the sweep — e.g.,
"diff bites" stems to {diff, bit} (not {diff, bite}) and lost the intended overlap with "bite me" queries; corrected to "diff bite". When adding new aliases later, verify each candidate with _tokenize() first.
- No removal or rewording of existing frontmatter aliases. Promotion is purely additive.
After promotion lands
- attune-help releases a new patch/minor with these aliases in frontmatter.
- attune-rag bumps its
attune-help>=X.Y.Z,<X.(Y+1) floor pin.
- attune-rag removes the corresponding entries from
aliases_override.json (the file can either go away entirely or stay as a stub for future temporary additions — design call at that point).
- attune-rag re-runs
docs/specs/embedding-retriever/run_diagnostic_1.py against the new attune-help to verify the paraphrased R@3 ≥ 70% result is preserved end-to-end.
Related
Context
attune-rag completed an alias-expansion-sweep (
docs/specs/alias-expansion-sweep/) that closed a 41pp R@3 gap on paraphrased developer queries against the bundled help corpus. The mechanism:KeywordRetrievercollapsing on paraphrased queries (P@1 97.5% → 11.25%, R@3 100% → 28.75%).These aliases currently live in
attune-rag/src/attune_rag/corpus/aliases_override.jsonas a temporary home — that override mechanism was built so the sweep could ship without requiring a new attune-help release per cluster. The long-term home is attune-help frontmatter so the aliases ship with the templates themselves.What this issue tracks
Promote the 89 proven aliases (across 6 templates) from attune-rag's override file into upstream attune-help frontmatter. One attune-help release; six file edits; no behavior change. After the upstream release lands and attune-rag pins to the new attune-help version, attune-rag's
aliases_override.jsonentries become redundant and can be deleted.Inventory — 89 aliases across 6 templates
Source: attune-rag/src/attune_rag/corpus/aliases_override.json
concepts/tool-bug-predict.md(20 aliases)Append to existing frontmatter
aliases:list:concepts/tool-security-audit.md(14 aliases)Frontmatter currently has empty
aliases: []. Add:concepts/tool-release-prep.md(14 aliases)Append to existing frontmatter
aliases:list:concepts/tool-smart-test.md(14 aliases)Frontmatter currently has empty
aliases: []. Add:concepts/tool-fix-test.md(13 aliases)Append to existing frontmatter
aliases:list:concepts/tool-code-quality.md(15 aliases)Frontmatter currently has empty
aliases: []. Add:Constraints to honor when promoting
DirectoryCorpus) raisesDuplicateAliasErrorif any alias appears on two templates. The 89 aliases above were chosen so they don't collide with each other or with existing frontmatter aliases as of attune-help 0.11.0. A pre-merge check should re-verify against whatever main is at promotion time.KeywordRetriever.MIN_ALIAS_OVERLAP = 2requires at least two query-token overlaps before crediting an alias. Single-token aliases get silently dropped. Every alias above was pre-validated withattune_rag.retrieval._tokenize."diff bites"stems to{diff, bit}(not{diff, bite}) and lost the intended overlap with"bite me"queries; corrected to"diff bite". When adding new aliases later, verify each candidate with_tokenize()first.After promotion lands
attune-help>=X.Y.Z,<X.(Y+1)floor pin.aliases_override.json(the file can either go away entirely or stay as a stub for future temporary additions — design call at that point).docs/specs/embedding-retriever/run_diagnostic_1.pyagainst the new attune-help to verify the paraphrased R@3 ≥ 70% result is preserved end-to-end.Related
tasks.mdM13.4 for the originating tracking entry.