Configuration

Environment variables

Scripts and adapters read configuration from environment variables. The recommended way to set them is a .env file in your working directory — it is loaded automatically via python-dotenv.

LLM access

schematize talks to your LLM through the OpenAI-compatible chat completions API (via ChatOpenAI), so any provider or server implementing that API works — the official OpenAI API, LiteLLM, vLLM, Ollama, and many others. Point API_URL at whichever one you're using.

Variable	Required	Description
`API_KEY`	Yes	API key for the LLM provider
`API_URL`	No	Base URL for the LLM endpoint. Omit for the provider default (the official OpenAI API). Set for LiteLLM, vLLM, Ollama, or any other OpenAI-compatible endpoint.

Example .env:

API_KEY=sk-...
API_URL=https://api.openai.com/v1

For a local Ollama model:

API_KEY=ollama
API_URL=http://localhost:11434/v1

For a self-hosted vLLM server:

API_KEY=vllm
API_URL=http://localhost:8000/v1

Using any provider via LiteLLM

Because API_URL accepts any OpenAI-compatible endpoint, the recommended way to reach any LLM is to run a LiteLLM proxy and point schematize at it — this is the setup we used for our experiments. Start the proxy, then:

API_KEY=sk-litellm
API_URL=http://localhost:4000

The model name (model_name for the CLI, or model= for ChatOpenAI) is routed by the proxy to the underlying provider — OpenAI, Anthropic, Gemini, local models, etc. — so you switch providers without touching schematize.

Weaviate

Required only when using the [weaviate] extra.

Variable	Default	Description
`WV_URL`	`localhost`	Weaviate host
`WV_PORT`	`8080`	Weaviate HTTP port
`WV_GRPC_PORT`	`50051`	Weaviate gRPC port
`WV_API_KEY`	(empty)	Weaviate API key

Example:

WV_URL=weaviate.example.com
WV_PORT=8080
WV_GRPC_PORT=50051
WV_API_KEY=my-weaviate-key

Prompts

Bundled prompts are YAML files under src/schematize/configs/prompt/ and are included as package data. Load them with:

from schematize import load_prompts

prompts = load_prompts(language="en", system_type="law")

`language`	`system_type`	Domain
`"en"`	`"law"`	English legal judgments
`"en"`	`"tax"`	English tax interpretations
`"en"`	`"general"`	Domain-neutral, for any document corpus
`"pl"`	`"law"`	Polish legal judgments
`"pl"`	`"tax"`	Polish tax interpretations

The paper's experiments use the pl/law combination; the others are additional prompt sets for use beyond the paper.

load_prompts returns a flat dict of named prompt strings. Wrap it in SchemaGeneratorPrompts before passing to SchemaGenerator:

from schematize import SchemaGenerator, SchemaGeneratorPrompts

generator = SchemaGenerator(llm=llm, retriever=retriever, prompts=SchemaGeneratorPrompts(**prompts))

To use custom prompts, replace individual keys in the dict before wrapping:

prompts = load_prompts(language="en", system_type="law")
prompts["schema_generator_prompt"] = "Your custom schema generator prompt..."

generator = SchemaGenerator(llm=llm, retriever=retriever, prompts=SchemaGeneratorPrompts(**prompts))

See Prompts for the directory convention, required placeholders per prompt, and how to bundle a new language or system_type/domain.

LangChain LLM cache

For development, enable caching to avoid redundant LLM calls:

from schematize.utils.langchain import setup_langchain_llm_cache

setup_langchain_llm_cache()  # uses SQLite by default

This calls langchain.globals.set_llm_cache with a SQLite-backed cache. It is not enabled by default in the library itself.