ratel-ai 0.6.0

Python SDK for Ratel — context engineering platform for AI agents. BM25 tool retrieval, MCP ingestion, framework-neutral capability tools.

ratel-ai

Context engineering for Python agents.

Docs • GitHub • Discord

ratel-ai retrieves the tools and skills relevant to each agent turn instead of sending the full catalog to the model. It bundles Ratel's Rust engine in-process: BM25 by default, with configurable semantic and hybrid retrieval available when needed. The default and local-model paths require no API key, vector database, or service. Installing a published package on a supported prebuilt target also requires no Rust toolchain.

Use ToolCatalog for ranked tools with sync or async handlers and SkillCatalog for ranked Markdown playbooks loaded on demand. Expose search_capabilities_tool, invoke_tool_tool, and get_skill_content_tool so an agent can discover tools and skills, invoke tools, and load full skill instructions. Tools from existing MCP servers can be ingested into the tool catalog with the mcp extra.

Semantic and hybrid retrieval use a configurable embedding model (ADR 0012), set per catalog via the embedding argument: the built-in default, a HuggingFace repo or local directory (in-process), or an OpenAI-compatible endpoint (OpenAI, Ollama, TEI, vLLM).

For semantic or hybrid retrieval, register() folds embedding in: it accepts one tool or a whole batch and embeds on a worker thread, so model loading, HTTP, and inference never block the asyncio loop or hold the GIL — and embedding errors surface right at register():

async def retrieve(tools):
    catalog = ToolCatalog(method="semantic", embedding={"ollama": "nomic-embed-text"})
    await catalog.register(tools)                              # embeds the batch here
    return await catalog.search_async("deploy the service", 5)

register() is async for every method (BM25 too); search() stays synchronous for BM25 only, and search_async() covers all three. To change the endpoint's model or vector dimension, construct a new catalog and re-register.

Install

pip install ratel-ai
# MCP ingestion: pip install 'ratel-ai[mcp]'

Quickstart

Save as quickstart.py, then run python quickstart.py:

import asyncio
from ratel_ai import ExecutableTool, ToolCatalog

async def main():
    catalog = ToolCatalog()
    await catalog.register(
        ExecutableTool(
            id="get_weather",
            name="get_weather",
            description="Get the current weather for a city.",
            input_schema={"properties": {"city": {"type": "string"}}},
            output_schema={"type": "object"},
            execute=lambda args: {"forecast": f"Sunny in {args['city']}"},
        )
    )

    hit = catalog.search("What is the weather in Rome?", 1)[0]
    print(await catalog.invoke(hit.tool_id, {"city": "Rome"}))


asyncio.run(main())

Continue with the Python guide, capability tools, API reference, or the Pydantic AI example.

Telemetry export is optional. With the otlp extra installed, configure_telemetry() reads RATEL_OTLP_ENDPOINT (falling back to the superseded RATEL_URL, which warns) and RATEL_API_KEY, wires trace and Logs exporters, and returns a shutdown handle. It exports only gen_ai.*/ratel.* signal spans and EventRecords by default — export_all_spans=True widens spans only. Message/tool content stays off by default; opt in with capture_content/include_span_and_events (see the telemetry guide for the capture modes and their privacy implications). Hosts that already own OpenTelemetry providers add both ratel_span_processor and ratel_log_record_processor instead.

Package layout: ratel_ai/ is the Python surface, native/ contains the PyO3 binding, and tests/ exercises both. For local development, create .venv with uv, install maturin, pytest, pytest-asyncio, ruff, and mypy, then run .venv/bin/maturin develop and .venv/bin/pytest.