sutram 0.1.2

A lightweight Python library for calling LLMs through a unified interface with built-in caching, retry policies, and session management

sūtram (సూత్రం)

The thread that connects

A unified Python interface for LLM providers. One thread to connect your code to any language model — with built-in caching, retry policies, and multi-turn session management.

In Sanskrit, sūtram means "thread" or "formula" — the essential connection that holds everything together.

Features

• Unified Provider Interface — One API to call OpenRouter, OpenAI, Anthropic, and more
• Built-in Caching — Avoid redundant API calls with pluggable cache backends
• Retry Policies — Configurable exponential/fixed backoff with status-code filtering
• Multi-turn Sessions — First-class support for conversation history management
• Sync & Async — Full support for both synchronous and asynchronous workflows
• Extensible — Add new providers by subclassing BaseProvider

Installation

pip install sutram

Quick Start

from sutram import create_provider, Session, DictCache

# Create a provider
provider = create_provider(
    name="openrouter",
    model="openai/gpt-4",
    api_key="your-api-key",
    cache=DictCache(),
)

# Single-turn call
response = provider.call_llm("What is the meaning of sūtram?")
print(response)

# Multi-turn conversation
session = Session(system_prompt="You are a helpful assistant.")
session.add_user_message("Hello!")
response = provider.chat(session.get_messages())
session.add_assistant_message(response)
session.add_user_message("Tell me more.")
response = provider.chat(session.get_messages())

Configuration

from sutram import create_provider, DictCache

provider = create_provider(
    name="openrouter",
    model="openai/gpt-4",
    api_key="your-api-key",
    max_retries=3,
    backoff_factor=1.0,
    strategy="exponential",
    timeout=120,
    retry_on_status=[429, 500, 502, 503, 504],
    cache=DictCache(),
)

Adding a Custom Provider

from sutram import BaseProvider, PROVIDER_REGISTRY

class MyProvider(BaseProvider):
    def _build_request_body(self, messages: list[dict]) -> dict:
        return {"model": self.model, "messages": messages}

    def _parse_response(self, data: dict) -> str:
        return data["choices"][0]["message"]["content"]

PROVIDER_REGISTRY["myprovider"] = {
    "cls": MyProvider,
    "base_url": "https://api.myprovider.com/v1/chat",
}

License

MIT