Why:
- Support Text PDFs in the same DocumentTree/markdown contract as DOCX.
Changes:
- PyMuPDF parser, text-layer gate, shared heading heuristics, upload dispatch for .pdf.
Impact:
- Scanned/image PDFs are rejected at upload; requires pymupdf installed.
Co-authored-by: Cursor <cursoragent@cursor.com>
Why:
- Swagger UI is not user-friendly for non-technical users
- Need a proper web app for document management, querying, benchmarking, and administration
Changes:
- Babel-in-browser React dashboard with 5 tabs (Home, Documents, Query, Benchmarks, Admin)
- Dark blue theme matching benchmark report pages (#0F1419, #1A2332, #F59E0B)
- Comparison view for strategy analysis across documents
- Expandable experiment details with delete
- Admin panel with collapsible sections (Health, Qdrant, Chunk Preview, Questions, Cost Estimator)
- Document upload, processing, and strategy selection
- Query interface with chunk inspection
Impact:
- Dashboard served at /app/ via FastAPI StaticFiles
- All Swagger functionality available through web UI
Why:
- Users need to see each strategy's overall score at a glance in the report
- Technical report should show average metrics across all strategies
Changes:
- Each strategy card now shows its overall score below the metric bars
- Technical report aggregate table includes an AVERAGE row across all strategies
Impact:
- Managerial report: strategy cards now show overall score
- Technical report: aggregate table has a new AVERAGE row at the bottom
Why:
- Experiment list and detail responses showed raw IDs instead of filenames
- No way to delete experiments from the UI
Changes:
- Add document_filename field to ExperimentDetailResponse
- Enrich list endpoint with filenames and calculated best_strategy from aggregate_metrics
- Add DELETE /experiments/{id} endpoint
Impact:
- API responses now include document_filename for all experiment endpoints
- Frontend can display filenames instead of IDs
Why:
- Dashboard needs system health, Qdrant management, chunk preview, questions, and cost estimation endpoints
Changes:
- Add admin router with 11 endpoints (health, Qdrant CRUD, chunk preview, questions management, cost estimation)
- Add delete_experiment to SQLite storage
- Mount admin router and dashboard static files at /app
Impact:
- New /admin/* API routes available
- Dashboard served at /app/ via StaticFiles
Why:
- Need visibility into API usage and performance
- Need to track request method, path, status, and duration
- Helps debugging production issues
Changes:
- Added RequestLoggingMiddleware class
- Logs at appropriate level (INFO/WARNING/ERROR)
- Non-invasive, no logic changes
Why:
- Need request/response models for benchmark endpoints
- Need routes for creating and retrieving benchmarks
- Need view parameter for managerial vs technical report views
Changes:
- models.py: Added BenchmarkRequest, BenchmarkResponse, StrategyMetrics, ExperimentDetailResponse
- routes.py: Added POST /benchmarks, GET /benchmarks/{id}, GET /experiments, view parameter for reports
Why:
- Need LLM-as-Judge evaluation for automated scoring
- Need benchmark orchestration to run questions × strategies
- Need HTML report generation with two views (managerial/technical)
Changes:
- evaluation.py: LLM-as-Judge scoring on 4 metrics (context, similarity, faithfulness, hallucination)
- benchmark_service.py: Orchestration with per-strategy failure isolation
- report.py: Dual-view HTML reports with dark mode, charts, and strategy cards
Why:
- Need structured logging for debugging query pipeline
- Need to mount benchmarking router
Changes:
- Added logging.basicConfig() with INFO level
- Mounted benchmarking router
- Registered QueryError exception handler
Why:
- Need to list documents to get document IDs for queries
- ProcessRequest should default to all 5 strategies
- paragraph_count should be optional (not always stored)
Changes:
- Added GET /documents endpoint with pagination
- Added DocumentListResponse model
- ProcessRequest now defaults to all 5 strategies
- DocumentResponse.paragraph_count now defaults to 0
Why:
- Qdrant was storing only metadata, not chunk text
- Queries returned empty text in retrieved chunks
- Scores were low (~0.25) because embeddings were on empty strings
Changes:
- Added 'text' field to ChunkMetadata model
- Updated chunk_to_metadata() to copy text
Why:
- App factory needs to mount document routes and initialize SQLite at startup
Changes:
- Import and mount documents_router
- Call init_db() on app creation
Why:
- Need abstract base class for all chunking strategies
- Need OpenAI embedding service (text-embedding-3-small)
- Need orchestration to run chunk → embed → store pipeline
Changes:
- Base: ChunkingStrategy ABC, token counting, chunk ID generation, sentence splitting
- Embedding: batch embedding with 2048 text limit per call
- Service: strategy registry, single/multi-strategy runners with per-strategy failure isolation
Why:
- Need to parse .docx/.doc files into DocumentTree + markdown
- Need API endpoints for document upload, processing, and deletion
Changes:
- Parser: .doc→.docx conversion via LibreOffice, XML-level extraction for table-heavy docs, heuristic heading detection
- Service: upload, process (delegates to chunking), delete orchestration
- Routes: POST /documents, POST /documents/{id}/process, DELETE /documents/{id}, GET /strategies
