Commit Graph

3 Commits

Author SHA1 Message Date
Woody f2115ae563 feat: structured LLM output for decompose + citation fuzzy matching (Phase 5)
Phase 5.1 — Structured LLM output for query decomposition:
- Add SubQuestions Pydantic model with sub_question, keywords, rationale
- Add LLMClient.complete_structured() using langchain with_structured_output
- Update QueryDecomposer with structured output path + legacy json.loads fallback
- Update SQLite seed templates: add subq+citation labeling requirement
- Add tests: structured output, subquestions model validation, logging

Phase 5.2 — Citation format alignment and fallback links:
- Add document_id to SourceMetadata (backend + frontend types)
- Rewrite citationParser.ts with fuzzy matching and fallback document links
- Add RAGDatabasePage auto-expand from ?document= URL param
- Tighten generate_per_subq seed prompt: 'Copy exact bracket labels shown'
- Add citation parser tests for fuzzy match and fallback link scenarios
- Defer: DOCX/TXT PDF generation → Phase 5.3 (fallback links sufficient)
2026-04-28 15:39:17 +08:00
Woody 7d072e5ea1 feat(frontend): redesign ResponsePanel for per-sub-question sections with grouped sources
Redesign ResponsePanel with SubQuestionSections component that parses answer markdown on ## Sub-question N: boundaries and renders per-sub-question cards with headers, ReactMarkdown body, and collapsible sources scoped to each section. Extract FlatResponse for backward compatibility when subQuestionSources is null. Add buildCitationLookupForSubq and processCitationsForSubq for per-sub-question citation lookup scope isolation. Add anchor links in ExtractedQuestionsDisplay that smooth-scroll to matching ResponsePanel sections. Pass subQuestionSources through LTTPage.

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
2026-04-26 23:29:51 +08:00
Woody 9095432806 feat(frontend): add citation parser utility with tests (sub-phase 2.6)
processCitations() parses [filename, page N] patterns from LLM answers.
Cross-references with sources[] array to build clickable markdown links.
Graceful fallback: unmatched citations remain as plain text.
Handles markdown images/links, case-insensitive matching, DOCX without pages.

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
2026-04-24 17:53:10 +08:00