Healthy Diet AI Agent is a backend service, written in TypeScript on Bun with Express, that answers nutrition questions, analyses food photos, and grounds its advice in ingested dietary guideline documents. The agent runtime is built on LangChain, LangGraph, and DeepAgents, calls models through an OpenAI-compatible API with optional Google Gemini routing, and is reachable over an HTTP API or a terminal CLI.
It runs in two modes. Standalone mode uses SQLite, bootstraps its schema on startup, and ships as the default Docker Compose configuration, so no external services are needed. Integration mode stores chat history, user profiles, and document metadata in Supabase and keeps the API routes of the original health-diet-api project, from which this service was split out.
Knowledge comes from uploaded documents and a local knowledge base: a PDF-to-Markdown tool converts official guideline PDFs into structured Markdown that keeps table content as prose, a Version-Aware and Policy-Aware RAG engine resolves conflicts between multiple versions of a dietary guideline, and a knowledge graph module extracts and searches structured relations between health and diet concepts. A background job syncs clarification and reference content from Taiwan's Ministry of Health and Welfare (MOHW) on a configurable interval. Agent behaviour defaults live in agent_config.json, and environment variables override them per deployment.
Features
- Nutrition chat: dietary advice, meal planning, and nutrition Q&A through an agent runtime
- Food image analysis: a workflow that interprets meal photos and connects them to nutrition guidance
- RAG document management: upload, parse, embed, and query nutrition documents
- Version-aware retrieval: parameterizable rules resolve temporal conflicts between guideline versions, with a frozen evaluation suite under experiments/
- PDF cleaning tool: converts guideline PDFs to Markdown with prose descriptions of tables
- Knowledge graph: extraction and search of structured health and diet relations
- MOHW sync: scheduled import of public clarification content, toggled and timed by environment variables
- Switchable storage: SQLite for standalone use or Supabase for integration with the original stack
- Two interfaces: HTTP API via Express and a CLI with configurable user and thread IDs
- Docker Compose: standalone SQLite mode by default
