Think

Think is a Python package for creating thinking programs.

It provides simple but powerful primitives for composable and robust integration of Large Language Models (LLMs) into your Python programs.

Think supports OpenAI, Anthropic, Google (Gemini), Amazon (Bedrock) and Groq as LLM providers, Ollama for local models, as well as any OpenAI-compatible LLM API provider.

Examples

Ask a question:

from asyncio import run

from think import LLM, ask

llm = LLM.from_url("anthropic:///claude-3-haiku-20240307")

async def haiku(topic):
    return await ask(llm, "Write a haiku about {{ topic }}", topic=topic)

print(run(haiku("computers")))

Get answers as structured data:

from asyncio import run

from think import LLM, LLMQuery

llm = LLM.from_url("openai:///gpt-4o-mini")

class CityInfo(LLMQuery):
    """
    Give me basic information about {{ city }}.
    """
    name: str
    country: str
    population: int
    latitude: float
    longitude: float

async def city_info(city):
    return await CityInfo.run(llm, city=city)

info = run(city_info("Paris"))
print(f"{info.name} is a city in {info.country} with {info.population} inhabitants.")

Integrate AI with custom tools:

from asyncio import run
from datetime import date

from think import LLM, Chat

llm = LLM.from_url("openai:///gpt-4o-mini")

def current_date() -> str:
    """
    Get the current date.

    :returns: current date in YYYY-MM-DD format
    """
    return date.today().isoformat()

async def days_to_xmas() -> str:
    chat = Chat("How many days are left until Christmas?")
    return await llm(chat, tools=[current_date])

print(run(days_to_xmas()))

Use vision (with models that support it):

from asyncio import run

from think import LLM, Chat

llm = LLM.from_url("openai:///gpt-4o-mini")

async def describe_image(path):
    image_data = open(path, "rb").read()
    chat = Chat().user("Describe the image in detail", images=[image_data])
    return await llm(chat)

print(run(describe_image("path/to/image.jpg")))

Use Pydantic or custom parsers for structured data:

from asyncio import run
from ast import parse

from think import LLM, Chat
from think.parser import CodeBlockParser
from think.prompt import JinjaStringTemplate

llm = LLM.from_url("openai:///gpt-4o-mini")

def parse_python(text):
    # extract code block from the text
    block_parser = CodeBlockParser()
    code = block_parser(text)
    # check if the code is valid Python syntax
    try:
        parse(code)
        return code
    except SyntaxError as err:
        raise ValueError(f"Invalid Python code: {err}") from err

async def generate_python_script(task):
    system = "You always output the requested code in a single Markdown code block"
    prompt = "Write a Python script for the following task: {{ task }}"
    tpl = JinjaStringTemplate()
    chat = Chat(system).user(tpl(prompt, task=task))
    return await llm(chat, parser=parse_python)

print(run(generate_python_script("sort a list of numbers")))

For detailed documentation on usage and all available features, please refer to the code docstrings and the integration tests.

Retrieval-Augmented Generation (RAG) support

Think supports RAG using TxtAI, ChromaDB and Pinecone vector databases, and provides scaffolding to integrate other RAG providers.

Example usage:

    from asyncio import run

    from think import LLM
    from think.rag.base import RAG, RagDocument

    llm = LLM.from_url("openai:///gpt-4o-mini")
    rag = RAG.for_provider("txtai")(llm)

    async def index_documents():
        data = [
            RagDocument(id="a", text="Titanic: A sweeping romantic epic"),
            RagDocument(id="b", text="The Godfather: A gripping mafia saga"),
            RagDocument(id="c", text="Forrest Gump: A heartwarming tale of a simple man"),
        ]
        await rag.add_documents(data)

    run(index_documents())
    query = "A movie about a ship that sinks"
    result = run(rag(query))
    print(result)

You can extend the specific RAG provider classes to add custom functionality, change LLM prompts, add reranking, etc.

RAG evaluation is supported via the think.rag.eval.RagEval class, supporting Context Precision, Context Recall, Faithfulness and Answer Relevance metrics.

Quickstart

Install via pip:

pip install think-llm

Note that the package name is think-llm, not think.

You'll also need to install the providers you want to use: openai, anthropic, google-generativeai, groq, or ollama.

You can set up your LLM credentials via environment variables, for example:

export OPENAI_API_KEY=<your-openai-key>
export ANTHROPIC_API_KEY=<your-anthropic-key>
...

Or pass them directly in the model URL:

from think import LLM

llm = LLM.from_url(f"openai://{YOUR_OPENAI_KEY}@/gpt-4o-mini")

In practice, you might want to store the entire model URL in the environment variable and just call LLM.from_url(os.environ["LLM_URL"]).

Model URL

Think uses a URL-like format to specify the model to use. The format is:

provider://[api_key@][host[:port]]/model[?query]

• provider is the model provider, e.g. openai or anthropic
• api-key is the API key for the model provider (optional if set via environment)
• server is the server to use, useful for local LLMs; for OpenAI and Anthropic it should be empty to use their default base URL
• model-name is the name of the model to use

Examples: - openai:///gpt-3.5-turbo (API key in environment) - openai://sk-my-openai-key@/gpt-3-5-turbo (explicit API key) - openai://localhost:1234/v1?model=llama-3.2-8b (custom server over HTTP) - openai+https://openrouter.ai/api/v1?model=llama-3.2-8b (custom server, HTTPS)

(Note that if the base URL is provided, the model must be passed as a query parameter.)

Using the URL format allows you to easily switch between different models and providers without changing your code, or using multiple models in the same program without hardcoding anything.

Roadmap

Features and capabilities that are planned for the near future:

• documentation
• support for local LLMs via HuggingFace
• more examples

If you want to help with any of these, please look at the open issues, join the conversation and submit a PR. Please read the Contributing section below.

Contributing

Contributions are welcome!

To ensure that your contribution is accepted, please follow these guidelines:

• open an issue to discuss your idea before you start working on it, or if there's already an issue for your idea, join the conversation there and explain how you plan to implement it
• make sure that your code is well documented (docstrings, type annotations, comments, etc.) and tested (test coverage should only go up)
• make sure that your code is formatted and type-checked with ruff (default settings)

Copyright

Copyright (C) 2023-2025. Senko Rasic and Think contributors. You may use and/or distribute this project under the terms of MIT license. See the LICENSE file for more details.