Godot's API moves between versions and models confidently produce methods that no longer exist; grounding answers in the docs you indexed is the fix, and because you build the index yourself it matches the version you actually use. The cost is the setup: four scripts run before the first question, and the collection name you produce has to be handed to the server at launch.
A Python server that serves Godot documentation to a model by semantic search over a Chroma collection you build from the official docs.
- Godot documentation retrieved by meaning, so a question phrased in your own words finds the right page
- A pipeline in the repository that builds the index: download the docs, convert reStructuredText to markdown, chunk it, and vectorize it
- A choice of embedding model — all-MiniLM-L6-v2 locally, or bge-m3 and bge-large-zh-v1.5 for stronger multilingual retrieval
- Everything served from a local Chroma database, so lookups do not leave your machine
No account and no key for the local path, but the index is not shipped: you run the download, convert, chunk and vectorize steps once before the server has anything to answer from. Python 3.12 with uv, and the server started with --chromadb-path and --collection-name pointing at what you built. Using an API-backed embedding model instead needs a key for that provider.