AI Cooking Assistant (LangGraph + Streamlit)

Generate structured recipes from natural language using a multi-agent LangGraph pipeline and a Streamlit UI.

Features

• Multi-agent pipeline: intent extraction, planning, ingredient optimization, steps, validation, confidence scoring.
• Streamlit UI for interactive use.
• CLI entrypoint for quick testing.
• YouTube transcript-based validation (optional, via yt-dlp + faster-whisper).

Setup

1. Create a virtual environment and install dependencies:

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Configure environment variables:

touch .env

Then add this line inside .env:

OPENROUTER_API_KEY="your_openrouter_key_here"

Run

Streamlit UI:

streamlit run streamlit_app.py

CLI:

python main.py

Docker

Build image:

docker build -t cooking-assistant .

Run container:

docker run --rm -p 8501:8501 --env-file .env cooking-assistant

Optional (CLI in container):

docker run --rm --env-file .env cooking-assistant python main.py

Architecture (LangGraph)

This project uses a LangGraph StateGraph with a shared state object that flows through specialized nodes.

State

Defined in state.py as CookingState. It is a Python dict that represents the current workflow data:

• Input: user_query
• Understanding: intent, constraints
• Planning: recipe_plan, optimized_ingredients, steps
• Validation and output: validation, final_recipe
• Meta: confidence, safety_flags, retries

Nodes

Defined in graph.py via graph.add_node(...). Each node reads state and returns an updated state:

• intent_extractor -> parse intent + constraints
• recipe_planner -> high-level plan
• ingredient_optimizer -> adapt ingredients
• step_generator -> generate steps
• recipe_validator -> validate and finalize
• confidence_scorer -> compute confidence score

Edges / Execution Order

The current graph is linear:

intent_extractor
  -> recipe_planner
  -> ingredient_optimizer
  -> step_generator
  -> recipe_validator
  -> confidence_scorer
  -> END

Graph definition lives in graph.py:

• Nodes: graph.add_node(...)
• Edges: graph.add_edge("from", "to")
• Entry point: graph.set_entry_point("intent_extractor")

Notes / Limitations

• The graph is currently sequential (no branching, retries, or feedback loops).
• YouTube validation depends on network access and can be slow or fail if YouTube changes its HTML.