README.md

Project Overview

AI Kitchen — a FastAPI service that acts as a "personal chef." Given a photo or list of ingredients, it uses an LLM agent to identify ingredients, search for recipes via Tavily, score them by nutrition and difficulty, and return structured meal recommendations.

Tech Stack

• Python >=3.12, managed with uv
• FastAPI for HTTP API
• LangChain (langchain, langchain-openai, langchain-tavily) for agent orchestration
• Qwen model (qwen3.5-flash) via DashScope, using OpenAI-compatible API
• Tavily for web search
• Alibaba Cloud OSS for object storage

Commands

# Install dependencies
uv sync

# Run the app
uv run uvicorn app.main:app --reload

Environment Variables

Copy .env.example to .env and set:

• DASHSCOPE_API_KEY — DashScope API key (Qwen model access)
• DASHSCOPE_BASE_URL — DashScope base URL (OpenAI-compatible endpoint)
• TAVILY_API_KEY — Tavily search API key

Architecture

app/
├── main.py              # FastAPI app entry point
├── agents/              # LangChain agent definitions
│   └── personal_chief.py  # Personal chef agent (ingredient → recipe recommendations)
├── api/v1/              # API route handlers
├── models/              # Pydantic models / data schemas
└── common/              # Shared utilities

Core flow: The agent in personal_chief.py loads a Qwen LLM model via init_chat_model (OpenAI-compatible provider), wires it with a TavilySearch tool, and provides a structured system prompt that enforces a 4-step pipeline: identify ingredients → search recipes → score by nutrition + difficulty → output structured report.