Installation¶

SchemaCast needs Python 3.11 or later and graflo 1.17 or later.

uv add schemacast        # or: pip install schemacast

To work on SchemaCast or run the examples, install from a clone:

git clone https://github.com/growgraph/SchemaCast
cd SchemaCast
uv sync --group dev

What it installs¶

Six runtime dependencies and nothing else: graflo, openai, pydantic, pydantic-settings, click and pyyaml. There is no LangChain, no encoder and no vector store; a test keeps it that way.

Command	What it is
`schemacast`	`sheet`, `select` and `cast`; see Command line
`schemacast-bench`	the benchmark

Pointing it at a model¶

Any OpenAI-compatible endpoint: OpenAI itself, Ollama's /v1, vLLM. Settings come from environment variables only; no file is read.

OpenAIOllamavLLM

export SCHEMACAST_LLM_API_KEY=sk-...
# SCHEMACAST_LLM_MODEL defaults to gpt-5.6-luna

export SCHEMACAST_LLM_BASE_URL=http://localhost:11434/v1
export SCHEMACAST_LLM_MODEL=qwen3:8b
export SCHEMACAST_LLM_API_KEY=local   # any value; without a key only cached replies are served

export SCHEMACAST_LLM_BASE_URL=http://localhost:8000/v1
export SCHEMACAST_LLM_MODEL=<served model name>
export SCHEMACAST_LLM_API_KEY=local

Without an API key SchemaCast makes no request at all, and serves only replies it has cached. Every setting is listed in Settings.