#cross-encoder

AI document assistant with a production RAG pipeline — hybrid BM25 + semantic search, cross-encoder reranking, and streaming answers. Built with FastAPI, Streamlit, ChromaDB, and DeepSeek LLM. Fully Dockerized and deployed on Railway.

A retrieval-based medical Q&A system that gives accurate answers to diabetes-related questions by searching a curated medical knowledge base instead of generating text from scratch. Built with a hybrid FAISS + BM25 retrieval pipeline, cross-encoder reranking, and multi-layer safety gating and served through a Streamlit web app.

多模态内容理解与工具编排 Agent：Chinese-CLIP 图文共享向量空间 + BM25/RRF 混合检索 + cross-encoder 重排（真实文本查询 hit@10 0.668→0.918）+ Schema 约束工具调用 + 可溯源生成，含 MCP、多轮会话记忆、ReAct 编排与分层离线评测。 | Multimodal agent: Chinese-CLIP retrieval, BM25+RRF hybrid search, cross-encoder reranking, schema-constrained tool calling, MCP server.

Production-grade AI Customer Support Agent for @AppleSupport using Kaggle Twitter dataset (~3M tweets). Features Hybrid RAG (ChromaDB + BM25 + RRF + Cross-Encoder), Risk Escalation Router, LLM-as-a-Judge Evaluation Harness, and FastAPI Web UI.