Spaces:
Build error
Build error
Commit History
-
fix: make ingestion and querying robust; add acceptance-test corpus
Shubham Pawar +1: Claude Sonnet 5Issues found while building an end-to-end test corpus and running it against the live API (see docs/fixes/2026-09-05-ingestion-and-event-loop-blocking.md): Ingestion - Markdown ingest returned HTTP 500 in offline / strict-TLS environments: UnstructuredMarkdownLoader downloads a spaCy model at runtime. Switched .md/.markdown to TextLoader (markdown is chunked as plain text anyway). - A single ingest/query froze the whole worker: async handlers ran the synchronous load->chunk->embed->store / retrieve->LLM work on the event loop. Moved it to run_in_threadpool; stream() driven via iterate_in_threadpool. /health now stays ~1ms responsive during a 5.5s ingest (was blocked ~4.75s). - /api/ingest had no upload limits. Added MAX_UPLOAD_FILES / MAX_FILE_SIZE_MB / MAX_TOTAL_UPLOAD_MB, enforced before any work, returning HTTP 413. - Embedding model now warmed in the lifespan startup hook instead of stalling the first request. Querying - Groq model llama-3.3-70b-versatile is decom…c8f7f2e
-
feat: implement dynamic retrieval for better context coverage
- Fetch 20 candidate chunks broadly (was fixed at top_k) - Filter by distance threshold (keep all with distance <= 1.5) - Cap at 15 chunks max to stay within LLM context limits - Log candidate vs relevant count for observability - Complex questions now get 3x more context than beforec4c601b -
fix: improve RAG retrieval quality for dense documents
- Increased chunk_size 1000 -> 1500 (prevents splitting relevant paragraphs) - Increased top_k 5 -> 8 (more context for large papers) - Rewrote prompt: synthesizes partial info instead of refusing - Prompt now instructs to use details, quotes, and structure answersc2702f6 -
fix: update Gemini model to gemini-2.5-flash in pipeline
- Synced pipeline model name with frontend config - Groq stays on llama-3.3-70b-versatile (confirmed stable)6d08e00
-
feat: add multi-provider LLM support (Gemini, OpenAI, Claude, Groq)
- Added provider selection cards on login page with tier badges - Dynamic Get Key links for each provider - Updated pipeline to support multiple LLM providers - Added langchain-openai, langchain-anthropic, langchain-groq deps - Updated API models and routes for provider parametera226579
-
feat: add multi-tenant BYOK authentication
- Added API key login page with validation - Added user namespace for collections (hash-based isolation) - Added collection list with create/delete in sidebar - Added logout functionality - Made API key required (removed server-side fallback) - Session isolation: each user's chats and collections are separate9a80835
-
feat: add BYOK (Bring Your Own Key) support
- Added api_key field to QueryRequest model - Updated query endpoint to pass user key to pipeline - Added _build_chain_with_key() for dynamic LLM creation - Added API key input field in sidebar with password mask - Users can now use their own Gemini API keyef93124
-
fix: fix linting issues
a6df84a
-
feat: add async/batch processing utilities
- Add AsyncBatchProcessor with concurrency control - Add AsyncDocumentProcessor for document pipelines - Add RateLimiter (token bucket algorithm) - Add AsyncRetry decorator with exponential backoff - Add gather_with_concurrency helper - Add run_async sync-async bridge - Add comprehensive unit tests (16 tests, all passing) - Add documentation in docs/explanations/phase5/dcb67f2 -
feat: add query caching with in-memory and Redis backends
- Add CacheEntry dataclass with TTL support - Add InMemoryCache with LRU eviction - Add RedisCache for distributed deployments - Add QueryCache with get_or_compute pattern - Add create_cache factory function - Add comprehensive unit tests (all passing) - Add documentation in docs/explanations/phase5/7dd3c91
-
feat: add cross-encoder and LLM reranking
98e014c
-
feat: integrate Langfuse observability
- Add LangfuseTracer class with span/generation support - Add LangChain callback handler integration - Add create_traced_chain for automatic tracing - Update RAG pipeline with conditional tracing895d205
-
feat: add end-to-end testing scripts
- Add comprehensive test_e2e.py test suite - Add interactive demo.py script - Add testing documentation7640b67 -
feat: implement basic retrieval pipeline
- Add RetrievalResult dataclass with context extraction - Add Retriever class with score filtering - Add RetrieverFactory for easy setup - Add create_retriever convenience function - Update package exportsca03d00
-
feat: integrate HuggingFace embeddings
- Add EmbeddingsManager class with lazy loading - Add 4 pre-configured embedding models - Add get_embeddings convenience function - Add list_available_models helper - Update package exportsdeb2f2c
-
feat: implement ChromaDB vector store
- Add VectorStore class with persistant storage - Add similarity search with/without scores - Add batch document processing - Add retriever integration for LangChain - Update package exports1ac24a4 -
feat: implement document chunking pipeline
- Add DocumentChunker class with 3 strategies (recursive, sementic, token) - Add chunk_documents convenience function - Include chunk metadata (index, size, strategy) - Update package exportsdd57ebc -
feat: implement multi-format document loaders
- Add MultiFormatDocumentLoader class supporting PDF, MD, TXT - Add load_document convenience function - Create sample test documents in data/sample_docs/d2f0e65
-
feat: initial project setup and structure
- Created project directory structure (src/, api/, rag/, llmops/, frontend/, evals/, tests/, config/, docs/) - Added README.md with project overview and quick start guide - Added .gitignore for Python, IDE, and environment files - Added .env.example with configuration template - Added requirements.txt with 2025-updated dependencies - Added LICENSE (MIT) - Added Python package __init__.py files - Added config/settings.py with Pydantic-based configurationa5f90d1