gluellm 1.1.4

A high-level Python SDK for Large Language Models with automatic tool execution, structured output support, multi-agent workflows, and evaluation data recording

GlueLLM

TL;DR: A high-level Python SDK for LLMs that handles the annoying stuff (tools, retries, structured output, batching) so you can ship features instead of glue code.

GlueLLM is opinionated in the “I’ve been burned by this in production” way. If you like sensible defaults, clear APIs, and fewer bespoke wrappers, you’ll feel at home.

Works great with Spiderweb

If you’re building RAG, you probably don’t just need LLM calls — you need crawling, extraction, chunking, validation, and storage too. That’s Spiderweb.

• GlueLLM: LLM calls + tool execution + structured output + embeddings + batching
• Spiderweb: documents/web → clean chunks → vector store → query

Tiny “together” example:

import asyncio
from gluellm import GlueLLM
from spiderweb import Spiderweb

async def main():
    async with Spiderweb(llm_client=GlueLLM()) as web:
        await web.crawl("https://example.com", ingest=True, save_to="./crawled")
        results = await web.query("What is this site about?", top_k=5)
        print(results.chunks[0]["content"][:200])

asyncio.run(main())

What is this?

GlueLLM is a high-level SDK that makes working with LLMs actually pleasant:

• You call complete() or structured_complete() and get results.
• Tools are plain Python functions.
• Retries and error classification are built-in.
• Batching and rate limiting are first-class.
• Providers are unified via any-llm-sdk.

Why you might like it

• Zero ceremony: minimal code to get real results
• Tool execution loop: automatic tool calling orchestration
• Structured output: Pydantic models, validated
• Provider-agnostic: one API for OpenAI, Anthropic, XAI, and others
• Embeddings: same ergonomics + error handling
• Batch processing: concurrency control, retry strategies, key pools
• Observability hooks: logging + optional tracing

Why you might not

• If you want a thin client that exposes every raw provider knob, GlueLLM isn’t trying to be that.
• If you hate opinions, you’ll hate opinions (mine included).

Installation

# Using uv (recommended)
uv pip install gluellm

# From source (dev)
uv pip install -e ".[dev]"

Quick start

Simple completion

import asyncio
from gluellm.api import complete

async def main():
    result = await complete(
        user_message="What is the capital of France?",
        system_prompt="You are a helpful geography assistant.",
    )
    print(result.final_response)

asyncio.run(main())

Tool calling (tools are just functions)

import asyncio
from gluellm.api import complete

def get_weather(location: str, unit: str = "celsius") -> str:
    """Get the current weather for a location."""
    return f"Weather in {location}: 22°{unit[0].upper()}, sunny"

async def main():
    result = await complete(
        user_message="What's the weather in Tokyo and Paris?",
        system_prompt="Use get_weather for weather queries.",
        tools=[get_weather],
    )
    print(result.final_response)

asyncio.run(main())

Structured output

import asyncio
from pydantic import BaseModel, Field
from typing import Annotated

from gluellm.api import structured_complete

class PersonInfo(BaseModel):
    name: Annotated[str, Field(description="Full name")]
    age: Annotated[int, Field(description="Age in years")]
    city: Annotated[str, Field(description="City of residence")]

async def main():
    person = await structured_complete(
        user_message="Extract info: John Smith, 35, lives in Seattle",
        response_format=PersonInfo,
    )
    print(person.model_dump())

asyncio.run(main())

Embeddings

import asyncio
from gluellm import embed

async def main():
    result = await embed("Hello, world!")
    print(result.dimension, result.tokens_used)

asyncio.run(main())

Configuration (the boring part)

Providers are configured via environment variables:

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
export XAI_API_KEY=xai-...

Models use provider:model strings:

• openai:gpt-4o-mini
• anthropic:claude-3-5-sonnet-20241022

Docs (when you want the details)

GlueLLM keeps deeper docs in docs/ so the README stays readable:

• docs/ARCHITECTURE.md
• docs/BATCH_PROCESSING.md
• docs/CONNECTION_POOLING.md
• docs/WORKFLOW_PATTERNS.md

More runnable examples live in examples/.

Contributing

PRs welcome. Please read CONTRIBUTING.md.

License

MIT — see LICENSE.