A kitchen-sink AI stack that actually cooks

What it does

CookHero is a full-stack diet and cooking assistant. It stores and searches recipes (including the entire HowToCook corpus), builds weekly meal plans, tracks what you actually ate via text or photo, and runs nutrition analysis against your goals. An LLM-powered agent ties it together with multi-turn chat, web search, and sub-agents for specialized tasks.

The interesting bit

The project treats “cooking assistant” as an excuse to implement the complete modern AI ops stack: hybrid vector+BM25 retrieval with reranking, Redis/Milvus dual caching, RAGAS evaluation with async quality scoring, token usage dashboards, prompt-injection guardrails, and even MCP server integration. The boring infrastructure is where the ambition lives.

Key highlights

• ReAct agent with pluggable sub-agents and an MCP tool hub; users can spin up custom experts from the UI
• Multi-modal intake: photo-based food logging, ingredient recognition, and DALL-E 3 recipe illustration
• Full observability layer: RAGAS faithfulness/relevancy metrics, per-request token tracing, LLM stats dashboard
• Security stack that would make a fintech blush: rate limiting, JWT lockouts, audit logs, and dual-layer prompt injection detection
• Ships with the entire HowToCook open-source recipe corpus pre-indexed

Caveats

• Heavy dependency footprint: PostgreSQL, Milvus, Redis, MinIO, Etcd, plus a half-dozen external API keys (Tavily, imgbb, vision models, reranker)
• README is upfront about being “面向厨房新手” (kitchen novices), but the infrastructure complexity suggests a hobbyist running a homelab rather than someone who struggles to boil water
• Some config sections in the README are truncated mid-file, so not all defaults are visible

Verdict

Worth a look if you’re building a reference architecture for a domain-specific AI assistant and want to see how the pieces fit together. Skip it if you just need a recipe app—Notion and a calorie tracker will load faster.