CookHero

CookHero is a full-stack personalized diet management platform built with FastAPI backend and React frontend, centered around an LLM Agent system that serves as an intelligent dietary assistant. The system employs a three-tier LLM routing strategy (fast/normal/vision), driving a Subagent expert system and a rich built-in toolset (diet plan management, nutrition analysis, RAG retrieval, Tavily web search, DALL-E 3 image generation, etc.) through ReAct reasoning loops, with MCP protocol support for custom tool extensions.

For knowledge retrieval, CookHero implements dual-path recall via vector search (Milvus) and BM25, combined with metadata filtering and Qwen3-Reranker-8B intelligent reranking, accelerated by Redis + Milvus dual-layer caching. Multi-modal capabilities support image recognition and joint text-image understanding, compatible with OpenAI vision model APIs (default Qwen/QVQ-72B-Preview), with a limit of 4 images per request at 10MB each. Business features cover weekly diet plan creation, AI-powered text/image dietary logging with automatic nutrition estimation, daily/weekly nutrition deviation analysis, and calorie/protein/fat/carb goal tracking.

Users can upload private Markdown recipes to build personalized knowledge bases that merge with the global HowToCook recipe library. The platform includes RAGAS-based RAG quality evaluation (Faithfulness, Answer Relevancy), LLM usage statistics, structured JSON audit logs (SIEM-compatible), and prompt injection protection via NeMo Guardrails and rule engines, with Redis sliding window rate limiting.

Deployment requires Python >= 3.12, Node.js >= 18, and recommends Docker Compose for infrastructure (PostgreSQL, Redis, Milvus, MinIO, Etcd). Licensed under Apache-2.0 with 421+ commits indicating active development. Recipe data is primarily Chinese; internationalization scope is to be confirmed. Vision models and Reranker depend on external APIs, limiting fully offline deployment.