wink-wink-wink555/PsyChat

Built on FastAPI with ChromaDB vector storage and DeepSeek LLM, the system implements a ReAct-based query optimization loop that refines search terms through multi-round iterations before retrieval. Beyond standard RAG, it reconstructs complete dialogue contexts from chunked data (eliminating fragmentation issues) and performs explicit counselor speech pattern analysis across four dimensions (empathy methods, questioning strategies, pacing, language style) with topic-level caching to minimize LLM calls. The architecture enforces continuous dialogue monitoring—triggering forced retrieval after N consecutive turns without RAG—and generates responses through a structured four-layer prompt that combines full case examples, matched dialogue fragments, extracted speech patterns, and planned interaction strategies.