strata-forge

Typed, async-first toolkit for AI/LLM work — inference, evaluation, fine-tuning, retrieval, and remote compute orchestration across the major foundation-model providers.

Built on LiteLLM for provider breadth and Pydantic v2 for strict typing. The public API is async, with sync wrappers in strata_forge.sync for CLI and notebook ergonomics. Heavy optional dependencies (training, serving, remote compute) live behind extras, so the base install stays light.

Install

pip install strata-forge

Requires Python 3.14+.

Optional extras pull in the heavy dependencies only when you need them:

pip install "strata-forge[rag]"        # embeddings, vector store, reranking
pip install "strata-forge[compute]"    # SkyPilot + SSH remote execution
pip install "strata-forge[serving]"    # vLLM
pip install "strata-forge[finetuning]" # torch, transformers, trl, peft, bitsandbytes
pip install "strata-forge[all]"        # everything

Full set: redis, multimodal, inspect, langfuse, hf, evals, rag, compute, serving, bedrock, storage, finetuning, all.

Quickstart

import asyncio

from strata_forge.llm import LLMClient, Message


async def main() -> None:
    client = LLMClient("gpt-5.5", provider="openai")
    response = await client.complete([Message.user("Say hello in one short sentence.")])
    print(response.text)


asyncio.run(main())

Provider credentials are read from the environment (for example OPENAI_API_KEY). A CLI is installed as strata-forge.

What's in it

Documentation

Reference docs live in the repository:

• Architecture overview
• Architecture Decision Records
• Per-module API reference

Development

uv sync        # materialize the venv + lockfile
make check     # lint (ruff) + type-check (pyright strict)
make test      # unit tests

License

Apache License 2.0.