Spaces:
Sleeping
Sleeping
Commit History
-
Fix model, chunking, and eval — 15/15 perfect score
- Switch to stepfun/step-3.5-flash:free (old Gemini model removed) - Fix chunking: split on H2/H3 only, merge H1 metadata into first section - Add retry logic for OpenRouter rate limits - Increase TOP_K to 7 for better retrieval coverage - Add eval results: 15/15 retrieval, 0 hallucination, all grounded - Fix encoding issues in eval script60f829a -
Complete RAG chatbot for Indecimal construction marketplace
- Core RAG pipeline: markdown-aware chunking, FAISS vector search, OpenRouter + Ollama LLM generation - Streamlit frontend with dark theme, chat UI, model comparison mode - 15-question automated evaluation script - 51 chunks indexed from 3 internal documents - Ready for HuggingFace Spaces deployment Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>565434f