keeptalking 0.3.0

LLM client library OpenAI should have made

keeptalking

A simple, pythonic interface to any OpenAI-compatible LLM server. You will never type response.choices[0].message.content ever again.

Installation

pip install keeptalking

Usage

The entire library is 3 functions:

from keeptalking import talk, write, vibe

Conversation

talk(model='google/gemini-2.5-flash', 
     roles=['system', 'user'], 
     messages=['Solve a math problem', 'Sum up all possible bases in which 97 is divisible by 17'],
     structure=int,
     tokens=10)

will use grammar constrained decoding and return a single integer with the answer. The return value of talk will always be of type structure, which defaults to str is omitted. If roles are omitted, the first message is considered a system message, the rest are user messages. If model is omitted, gemini-2.5-flash is used (default model can be overriden by setting the MODEL environment variable). If tokens is omitted, generation is limited to 2048 new tokens (default token limit can be overridden by setting the TOKENS environment variable).

The only parameter that should not be omitted is messages:

talk(['Solve a math problem. Provide your reasoning', 'Sum up all possible bases in which 97 is divisible by 17'])

write is an asynchronous version of talk that lets you beautifully parallelize batch requests:

asyncio.gather(
    write(model='google/gemini-2.5-flash', 
          roles=['system', 'user'], 
          messages=[sys, berry],
          structure=int,
          tokens=10)
    for berry in ['Strawberry', 'Blackberry', 'Raspberry', 'Blueberry', 'Canterbury']
)

write automatically self-throttles as necessary so it's safe to call thousands of write()s in parallel with no external rate limiting.

Vibe functions

Vibe functions are functions defined in natural language.

@vibe(model='google/gemini-2.5-flash', tokens=10)
def do_job(job_details):
    """System message"""
    return f"User message with {job_details}"

ELL users will notice that this format is shamelessly stolen inspired by ELL. However, keeptalking is much simpler than ELL. Despite being much simpler, keeptalking supports additional features like async vibe functions

@vibe()
async def homework_assistant(topic, pages=5):
    """Help the student with their homework"""
    return f"Write a {pages}-page essay on {topic}"

fully parallelizable like so

asyncio.gather(
    homework_assistant('Math'),
    homework_assistant('History'),
    homework_assistant('English')
)

and enabling structured outputs with a single type hint:

@vibe()
def count_rs(request) -> int:
    """Count how many Rs are in the request"""
    return request

unlike in the rest of the Python ecosystem, in vibe functions type hints actually ensure that the return value is always of the type in question

Backend configuration

The model server to be used is defined via environment variables. It can be defined directly by setting BASE_URL and API_KEY. If those are not set, keeptalking will default to OpenRouter if OPENROUTER_API_KEY is set, then OpenAI if OPENAI_API_KEY is set. Perv advanced users can monkey patch keeptalking.client_sync and keeptalking.client_async instead.

Example

You will find a detailed example in example.py. It takes top 10 models from openrouter's model catalog and reads text description of each model to filter out specialized models like coding or edit models, then runs a small test on each to check if they are working.