feat: lexical groups (per-namespace morph + abbrev)#62
Merged
Conversation
… UI, bindings
Per-namespace, per-language token normalization that runs at tokenize
time (both index-time on seeds and query-time on resolves). Two kinds:
morph (inflection variants of one root) and abbrev (short forms of a
phrase). Distinct from synonyms by design — only collapses tokens with
the same surface meaning, so it doesn't pollute sibling intents the way
the old L1 graph did.
Library
- microresolve::{LexicalGroup, LexicalKind} re-exported from crate root
- NamespaceHandle: list_lexical_groups, add_, remove_, update_
- src/lexical.rs (LexicalIndex, normalize_in_place) wired into all
tokenize call-sites in scoring + resolver_*
Server
- GET/POST /api/lexical-groups
- DELETE/PATCH /api/lexical-groups/{idx}
- POST /api/lexical-groups/suggest (operator-triggered LLM proposals
grounded in namespace vocab + intent descriptions; nothing applies
until approved)
- Every mutation lands in the per-key audit chain
(lexical_group.add / .remove / .update)
Studio UI
- New "Lexicon" page under Build with Inflections / Abbreviations tabs
- Manual add form + LLM Suggest panel with per-proposal approve/reject
- /lexical route + nav entry; existing /intents page untouched
Python bindings
- LexicalGroup pyclass + 4 namespace methods on Namespace
Node bindings
- LexicalGroup interface + 4 namespace methods (camelCase per napi);
index.d.ts regenerated
Persistence
- Resolver.lexical_groups is the source of truth, persisted in
_ns.json. Save now reads existing _ns.json first to preserve unknown
fields (compliance_frameworks, policy_overrides) so save/load is
round-trip safe across config the engine doesn't actively model.
EU AI Act pack
- Post-Omnibus refresh pulled forward from the tracing-explore branch:
6 new intent JSONs (ai_generated_csam, legitimate_use, ncii_adult,
realtime_remote_biometric_id, untargeted_facial_scraping) + 6
refreshed (biometric_categorization, emotion_recognition_workplace,
exploitation_vulnerability, predictive_policing, social_scoring,
subliminal_manipulation)
- Pack ships 10 morph groups + 3 abbreviations (rbi, ncii, csam)
- benchmarks/eu_ai_act_eval.py: 100 prohibited + 80 benign harness
(50 generic + 30 adjacent-legal carved out by Feb 2025 Commission
guidelines)
Empirical (head-to-head re-run, lexical_groups removed vs present):
thr=1.5 baseline: P=0.779 R=0.727 F1=0.742
thr=1.5 with lex: P=0.787 R=0.764 F1=0.766 (+2.3pp F1, +3.7pp R)
Improvement holds across the full threshold sweep (0.8 → 2.5).
Zero regression on CLINC150 + BANKING77.
Other
- 'status: experimental' tag on hipaa-triage, mcp-tools-generic,
safety-filter pack metadata
- .gitignore: ignore _skills/ (Claude Code local skills)
- benchmarks/test_lexical_groups.py: smoke test (4 query pairs)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gladius
force-pushed
the
feat/lexical-groups
branch
from
f388c8e to
2352f0d
Compare
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.
Per-namespace, per-language token normalization that runs at tokenize time.
Two kinds: morph (inflection variants of one root) and abbrev (short forms of a phrase). Distinct from synonyms — only collapses tokens with the same surface meaning, so it doesn't pollute sibling intents the way the old L1 graph did.
Surface
microresolve::{LexicalGroup, LexicalKind}re-exported from crate root.NamespaceHandle::{list, add, remove, update}_lexical_group.GET/POST /api/lexical-groups,DELETE/PATCH /api/lexical-groups/{idx},POST /api/lexical-groups/suggest(LLM-proposed, operator-approved). Audit-chain integrated.LexicalGroupclass + 4 namespace methods.LexicalGroupinterface + 4 namespace methods (camelCase),index.d.tsregenerated.Validation
Head-to-head ablation on EU AI Act pack (100 prohibited + 80 benign, 50 generic + 30 adjacent-legal carved out by Feb 2025 Commission guidelines):
Improvement holds across full threshold sweep (0.8 → 2.5). Zero regression on CLINC150 + BANKING77. EU AI Act pack ships 10 morph groups + 3 abbreviations (rbi, ncii, csam).
Also lands
benchmarks/eu_ai_act_eval.pyablation harnessbenchmarks/test_lexical_groups.pysmoke teststatus: experimentaltag on hipaa-triage / mcp-tools-generic / safety-filter🤖 Generated with Claude Code