Skip to content

feat(knowledge): retrieve cross-library evidence in one call - #578

Closed
LikiosSedo wants to merge 3 commits into
mainfrom
feat/knowledge-evidence-lookup
Closed

LikiosSedo wants to merge 3 commits into
mainfrom
feat/knowledge-evidence-lookup

Conversation

@LikiosSedo

@LikiosSedo LikiosSedo commented Sep 10, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Label routing misses facts that appear only in compiled Wiki bodies and requires separate reads before answering. Add knowledge_lookup to search mounted libraries and return bounded complete pages with original paths, library IDs and versions, registering only the returned evidence with existing citation validation.

The same lazy in-memory FTS5 engine powers siclaw knowledge search. Catalog navigation and library-aware label search remain available. The branch is integrated with the current CLI and capability contracts. No embedding service, production dependency, mount schema or deployment variable is added.

An offline corpus comparison command accepts original questions, supplied alternate queries and multiple expected pages. It compares label navigation, body lookup and equal-weight query fusion, reporting page recall and complete-question coverage. Supplied alternates are an evaluation intervention; runtime ranking remains unchanged. Documentation explains distractor-library, collection-order, applicability and no-answer evaluation boundaries.

Validation

  • Full backend suite: 379 files, 7,621 tests passed with the current pinned SDK installation patch.
  • Backend build, TypeScript no-emit check and AgentBox TypeScript check.
  • Focused retrieval, citation, capability, CLI snapshot and handoff contracts.
  • Portal capability catalog tests with an isolated configuration.
  • CLI help and deterministic synthetic retrieval comparison.
  • Corpus comparison smoke, known-baseline reconciliation and invalid-input rejection.

Remaining acceptance

This remains an experimental retrieval implementation. Representative multi-library Agent evaluation must establish normal tool selection, complete evidence coverage, applicable conditions, answer and citation correctness, and end-to-end latency. Lexical matches and relevance ranks are not answer confidence. Large pages still require separate reading.

Downstream analytics that count only native read inputs must recognize complete knowledge_lookup results before broad rollout. Private corpus questions, traces and measurements are retained separately.

Add an offline original-query and alternate-query comparison with page and complete-question coverage. Align lookup with the current CLI-only tool contract after rebasing.
@LikiosSedo
LikiosSedo force-pushed the feat/knowledge-evidence-lookup branch from c0824d1 to db019d0 Compare September 14, 2026 07:22
@LikiosSedo

Copy link
Copy Markdown
Collaborator Author

Superseded. Siclaw delivery has moved off this public repository; this change is continuing on the internal Siclaw repo. Closing the public PR.

@LikiosSedo LikiosSedo closed this Sep 15, 2026
@jacoblee-io
jacoblee-io deleted the feat/knowledge-evidence-lookup branch September 25, 2026 02:59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant