GrapeCity-AI/gc-qa-rag

Leverages a two-stage memory-focused approach for QA generation that dynamically adapts to document length—short documents use sentence-level precision, while long documents employ a "remember-then-focus" dialogue pattern to capture comprehensive coverage without hallucination. Beyond core QA pairs, it generates summaries, expanded answers, and question variants stored in a vector database (Qdrant), enhancing retrieval diversity and multi-turn dialogue capabilities. Built as a modular ETL-Retrieval-Generation stack with production-grade orchestration supporting Docker deployment, hybrid search with RRF ranking, and integration with major LLM APIs (OpenAI, Alibaba Bailian, etc.).