pydantic-ai-gigachat 0.1.0

PydanticAI integration for GigaChat

pydantic-ai-gigachat

PydanticAI integration for GigaChat.

This package implements a native PydanticAI Model against GigaChat's own /chat/completions API (not the OpenAI-compatibility shim), so all GigaChat features — Russian profanity filter, repetition penalty, function calling, streaming — are first-class.

Install

pip install pydantic-ai-gigachat

Requires Python 3.10+ and pydantic-ai-slim>=1.90.

Quick start

import asyncio
from pydantic_ai import Agent
from pydantic_ai_gigachat import GigaChatModel, GigaChatProvider

async def main():
    provider = GigaChatProvider()  # reads GIGACHAT_CREDENTIALS env
    model = GigaChatModel("GigaChat", provider=provider)
    agent = Agent(model, system_prompt="Отвечай коротко.")

    result = await agent.run("Привет, как дела?")
    print(result.output)

    await provider.aclose()

asyncio.run(main())

Configuration

Credentials

Get your Authorization key from the Sber developer cabinet (it's the base64-encoded client_id:client_secret). Pass it explicitly or via env:

export GIGACHAT_CREDENTIALS="<authorization key>"

provider = GigaChatProvider(
    authorization_key="...",
    scope=GigaChatScope.PERS,        # PERS / B2B / CORP
)

TLS certificates

GigaChat endpoints serve certificates signed by the Russian Trusted Root CA, which is not in the standard certifi bundle. Three options:

1. Production — download the CA from gu-st.ru and point the provider at it:

   export GIGACHAT_CA_BUNDLE_FILE=/path/to/russian_trusted_root_ca.cer
   

2. Append to certifi (one-shot, affects the whole venv):

   curl -k "https://gu-st.ru/content/Other/doc/russian_trusted_root_ca.cer" \
       >> $(python -m certifi)
   

3. Dev only — GigaChatProvider(verify_ssl=False).

Features

Feature	Status
Chat completions	✅
Streaming (SSE)	✅
Function calling	✅
OAuth2 token cache + refresh	✅
Custom CA bundle	✅
Pre-flight token counting (Model.count_tokens)	✅
Prompt cache reporting (cache_read_tokens)	✅
Embeddings	⏭ planned
/files (vision/RAG inputs)	⏭ planned

Available models

GigaChat, GigaChat-Pro, GigaChat-Max, GigaChat-2, GigaChat-2-Pro, GigaChat-2-Max (and -preview early-access variants).

model = GigaChatModel("GigaChat-Max", provider=provider)

Settings passthrough

PydanticAI's ModelSettings map cleanly to GigaChat parameters: temperature, top_p, max_tokens, n. GigaChat-specific knobs (repetition_penalty, update_interval, profanity_check) work via extra_body:

result = await agent.run(
    "...",
    model_settings={
        "temperature": 0.3,
        "max_tokens": 256,
        "extra_body": {"profanity_check": True, "repetition_penalty": 1.1},
    },
)

Examples

See examples/:

• basic.py — minimal agent
• tools.py — function calling
• streaming.py — streamed response

Development

python -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest

License

MIT.