This is an experiment to use LLM-based tools with structured output to migrate recipes from freeform text document to valid cooklang files.
Deploys a compose stack with a webserver and a vllm service. Drag and drop text files to convert and the server will convert them to cooklang recipes.
- 12GB VRAM
- docker
- Nvidia GPU
- just
- nix
- direnv
- Start the inference server with
just deploy(ordocker compose up). - Start the server with
just serve(or the corresponding command in justfile) - In your browser, go to
http://localhost:1337 - Drag and drop files.
- click convert.
The repo defines a GNBF grammar for cooklang (see ./assets).
Qwen3-8B-AWQ is run with reasoning + structured outputs to enforce the responses to conform to the cooklang grammar.
A 8B model does not seem powerful enough to use with grammars, it is unable to generate valid tokens and results in very long blocking time and truncated files.