acty-gigachat

acty-gigachat

acty-gigachat provides a GigaChatExecutor implementation for Acty. It adapts langchain-gigachat chat models to the Acty executor interface and adds session-aware behavior for GigaChat prefix-caching workflows.

Install

pip install acty-gigachat

For local development:

pip install -e .[dev]

Usage

import asyncio

from acty import ActyEngine, EngineConfig
from acty_gigachat import GigaChatExecutor
from langchain_gigachat.chat_models import GigaChat


async def main() -> None:
    model = GigaChat(credentials="YOUR_AUTHORIZATION_KEY", verify_ssl_certs=False)
    engine = ActyEngine(
        executor=GigaChatExecutor(model=model),
        config=EngineConfig(primer_workers=1, follower_workers=1),
    )
    try:
        submission = await engine.submit_group(
            "demo",
            {"input": "Hello!", "invoke_kwargs": {"temperature": 0.2}},
            [],
        )
        if submission.primer is not None:
            result = await submission.primer
            print(result.output)
    finally:
        await engine.close()


asyncio.run(main())

Payloads may provide messages, input, or prompt. The executor forwards an optional invoke_kwargs dictionary to the model.

Session Handling

GigaChatExecutor uses the GigaChat SDK session_id_cvar context variable to control the X-Session-ID header during model calls.

• if Job.context["cache"]["provider_ref"] is present, the executor reuses it
• otherwise, it creates a per-group fallback session id
• session-invalid failures rotate to a new session id
• transient failures reuse the current session id

This makes the executor suitable for Acty cache flows where a primer attaches a provider reference and followers reuse it later.

Telemetry

The executor can enrich the active OpenTelemetry/OpenInference LLM span with:

• Acty correlation identifiers
• bounded input message attributes
• cached prompt token usage when the model response exposes it

This is enabled by default and respects the configured telemetry privacy gates.

Development

• tests live under tests/
• the repo includes unit tests, session lifecycle tests, and Acty engine integration coverage for telemetry, retry, and TTL/session rotation behavior
• the package depends directly on both acty and acty-core because it imports symbols from both at runtime