​
Logging LLM requests

Get insight into individual LLM request, including prompt templates, input, output, duration, token usage cost, etc. Collect datasets for continuous iterations (fine-tuning, testing).

An LLM request represents a single query to an LLM provider. Baserun refers to the returned object as a Completion. If the request is successful, Baserun logs the completion in the UI as shown above.
A completion includes the input and output of the request, along with metadata such as the user, request ID, and model configurations. If the request fails, Baserun logs the error code and message in the LLM requests table.

​
Arguments

Beyond the default settings, users can customize additional arguments to improve analysis and data collection:

Specify the following properties as keyword arguments when creating an LLM request (e.g., using completions.create). These can also be adjusted on the completions object returned by the client.

client.chat.completions.create(
    name: Optional[str],
    user: Optional[str],
    session: Optional[str],
    metadata: Optional[Dict]
) -> Completion:

These arguments also apply to other completion-generating functions such as stream.

​
Instructions

1

Install Baserun SDK

pip install baserun
2

Set the Baserun API key

Create an account at https://app.baserun.ai/sign-up. Then generate an API key for your project in the settings tab. Set it as an environment variable:

export BASERUN_API_KEY="your_api_key_here"
3

Import and Init

In order to have Baserun trace your LLM Requests, all you need to do is import OpenAI from baserun instead of openAI. Creating an OpenAI client object automatically starts the trace, and all future LLM requests made with this client object will be captured.

from baserun import OpenAI


def example():
    client = OpenAI()
    completion = client.chat.completions.create(
        name="Paris Activities",
        model="gpt-4o",
        temperature=0.7,
        messages=[
            {
                "role": "user",
                "content": "What are three activities to do in Paris?"
            }
        ],
    )


if __name__ == "__main__":
    print(example())
4

Alternate init method

If you don’t wish to use Baserun’s OpenAI client, you can simply wrap your normal OpenAI client using init.

from baserun import init

client = init(OpenAI())
completion = client.chat.completions.create(
    ...
)

​
Tracing end-to-end pipelines

A Trace comprises a series of events executed within an LLM chain (also called a workflow, among other names). Tracing enables Baserun to capture and display the LLM chain’s entire lifecycle, whether synchronous or asynchronous.

Using Baserun, traces are tied to the client object of the library you are using. For example, if you are using the OpenAI library, you would create an OpenAI client object imported from baserun. When that client object is used all completions are automatically traced.

​
Arguments

These arguments can be passed when instantiating your client object, or can be set after instantiation.

OpenAI(name: str, **kwargs) -> OpenAI:

​
Instructions

In the following example, this pipeline has two LLM calls. Create a client at the beginning of the function you want to trace, and pass the client anywhere you want to use the same trace.

from baserun import OpenAI

def get_activities(client: OpenAI):
    response = client.chat.completions.create(
        model="gpt-4o",
        temperature=0.7,
        messages=[
            {
                "role": "user",
                "content": "What are three activities to do on the Moon?"
            }
        ],
    )
    return response.choices[0].message

def find_best_activity():
    client = OpenAI()
    client.name = "find_best_activity"
    client.user = "user123"
    client.session = "session123"

    moon_activities = get_activities(client)
    response = client.chat.completions.create(
        model="gpt-4o",
        temperature=0.7,
        messages=[
            {
                "role": "user",
                "content": "Pick the best activity to do on the moon from the following, including a convincing reason to do so.\n + {moon_activities}"
            }
        ],
    )
    client.result = "success"
    return response.choices[0].message

Alternatively, you can associate two events with the same trace or resume a trace using the trace_id. If you wish to associate an LLM request with a trace after the trace has completed, see the example below. Another common use case is when you want to add user feedback or tags to a trace after the pipeline has finished executing.

from baserun import OpenAI

def main():
    main_trace_id = str(uuid4())
    activities = get_activities(main_trace_id)
    find_best_activity(main_trace_id, activities)


def get_activities(trace_id: str) -> str:
    client = OpenAI(trace_id=trace_id)
    response = client.chat.completions.create(
        model="gpt-4o",
        temperature=0.7,
        messages=[
            {
                "role": "user",
                "content": "What are three activities to do on the Moon?"
            }
        ],
    )
    return response.choices[0].message

def find_best_activity(trace_id: str, activities: str) -> str:
    client = OpenAI(trace_id=trace_id)
    client.name = "find_best_activity"
    response = client.chat.completions.create(
        model="gpt-4o",
        temperature=0.7,
        messages=[
            {
                "role": "user",
                "content": "Pick the best activity to do on the moon from the following, including a convincing reason to do so.\n + {activities}"
            }
        ],
    )
    client.result = "success"
    return response.choices[0].message

​
Supported Models

At the moment, the Baserun Python SDK 2.0 supports all models using OpenAI library and Anthropic library, regardless of the underlying model’s provider. We are continuously adding support for new models. If you have a specific model you would like to use, please reach out to us at hello@baserun.ai or join our community.

If you use another provider or library, you can still use Baserun by manually creating “generic” objects. Notably, generic completions must be submitted explicitly using submit_to_baserun(). Here’s what that looks like:

from baserun.wrappers.generic import (
    GenericChoice,
    GenericClient,
    GenericCompletion,
    GenericCompletionMessage,
    GenericInputMessage,
)

question = "What is the capital of the US?"
response = call_my_custom_model(question)

client = GenericClient(name="My Traced Client")
completion = GenericCompletion(
    model="my custom model",
    name="My Completion",
    input_messages=[GenericInputMessage(content=question, role="user")],
    choices=[GenericChoice(message=GenericCompletionMessage(content=response))],
    client=client,
    trace_id=client.trace_id,
)
completion.submit_to_baserun()
Compare prompt versionsSessions