Export dataset run table

[marcklingen]

Describe the feature or potential improvement

For some use cases, it is useful to download the dataset run table as a CSV file. This allows to share it as a spreadsheet with non-technical teams who might not use Langfuse.

Additional information

No response

[bewestphal]

+1, Our enterprise has several non-technical folks interested in seeing the results of evaluations. Product, program management, legal to name a few.

[bewestphal]

The langfuse SDK with get_dataset_run was simple enough to make a script.

It's likely the contents of the CSV are unique to the contents of your traces anyway. A really simple script would look like:

def langfuse_run_to_csv(
    dataset_name: str,
    run_id: str,
    output_path: str,
):
    """
    Export langfuse run results as a CSV file. For historical runs.

    This function assumes the trace has only one observation for single user chat input

    Run with the following command:
    python -m {python_file_name} --dataset_name <dataset_name> --run_name <run_name> --output_path <output_path>
    """
    langfuse_client = Langfuse() # add your keys and stuff here
    dataset_run = langfuse_client.get_dataset_run(
        dataset_name=dataset_name, dataset_run_name=run_id
    )
    run_items = []
    for run_item in dataset_run.dataset_run_items:
        run_item_dict = {
            "id": run_item.id,
            "dataset_item_id": run_item.dataset_item_id,
            "trace_id": run_item.trace_id,
            "created_at": run_item.created_at,
            "updated_at": run_item.updated_at,
        }

        trace = langfuse_client.get_trace(id=run_item.trace_id)
        # We assume that the trace has only one observation for single user chat input
        observation = trace.observations[0]
        run_item_dict["trace_name"] = observation.name
        # Example - Get fields from json input
        run_item_dict["input"] = observation.input[0]["content"]

        # Example - Get fields from json output
        output = json.loads(observation.output)
        run_item_dict["response"] = output["response"] # this example is custom to my use case
        run_items.append(run_item_dict)

    df = pd.DataFrame(run_items)
    df.to_csv(output_path, index=False)


if __name__ == "__main__":
    args = argparse.ArgumentParser()
    args.add_argument("--dataset_name", type=str, required=True)
    args.add_argument("--run_name", type=str, required=True)
    args.add_argument("--output_path", type=str, required=True)
    args = args.parse_args()

    langfuse_run_to_csv(args.dataset_name, args.run_name, args.output_path)

Probably a better way. But this is only took ~30 minutes to write and observe my data in google sheets :)

[EyaChenenaoui]

Hey, are you sure this is script still works ?

[rekram1-node]

Would be super useful

[KokaChernov]

+1, really helpful feature once implemented

[MetcalfeTom]

We'd love this. Other output formats like markdown would also be very helpful for non-technical staff to read the outputs easily

[jannikmaierhoefer]

Hi all, we are currently working on enabling exports on this in the UI. In the meantime, you can find a guide on fetching the experiment results here: https://langfuse.com/faq/all/retrieve-experiment-scores