Spaces:
Runtime error
Runtime error
Commit History
-
Polish: BGE-large embeddings, contextual retrieval, 142 tests passing, lint clean
- Switch embedding model to BAAI/bge-large-en-v1.5 (MTEB top-ranked, 1024-dim) - Enable contextual retrieval (Anthropic Nov 2024) — 49% fewer retrieval failures - Enable HyDE, multi-query expansion, hybrid BM25+RRF, cross-encoder reranking - Fix FastAPI Path shadowing bug in api.py - Fix 3 failing tests → 142/142 passing - Fix ruff.toml format (was using pyproject.toml syntax) - Auto-fix + manually resolve all 257 lint issues → 0 errors - Add docs/rag_techniques.md and docs/system_architecture.md (demo ingestion content) - Update README with production config section and stack table - Re-ingest ChromaDB with new embedding model (BGE-large, incompatible with old MiniLM)21ca2ea
-
Security, bug fixes, and UX improvements
Joshua Sears +1: Claude Sonnet 4.6Security: - Strip API keys at config load time (field_validator) — single source of truth - Restrict CORS: remove allow_credentials=True, limit methods/headers - Block path traversal in /ingest: validate file_path stays within working dir - Hide internal paths from traceback exposure — log server-side, collapse UI Bugs: - Temp file always cleaned in finally block (was orphaned on exception) - Fix feedback export: replace deprecated mktemp() with NamedTemporaryFile + finally - Remove dead "web search fallback" UI message Performance: - Double-checked locking for cross-encoder singleton (thread safe) UX: - Download answer button on Hybrid RAG results - File uploader help text: size limit and scanned PDF warning - Actionable messages when context is borderline Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>076806c
-
Add A-RAG — hierarchical retrieval interfaces (Feb 2026)
Joshua Sears8316959 -
Add Speculative RAG — Google Research 2024 (~51% latency reduction)
Joshua Sears1f600ea -
Add TTRAG — test-time compute scaling for RAG (ICLR 2025)
Joshua Sears8f98b8a
-
Initial commit — production RAG system
Joshua Sears +1: Claude Sonnet 4.620+ retrieval techniques from 2022–2025 research: hybrid dense+BM25+RRF, cross-encoder reranking, CRAG, RAPTOR, GraphRAG, LightRAG, CoT-RAG, Adaptive RAG, Sufficient Context abstention, Agentic RAG, and more. FastAPI (35+ endpoints), Streamlit demo, Docker, CI/CD quality gate. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>bd8abf4