evaluate_comparative#
- langsmith.evaluation._runner.evaluate_comparative(experiments: Tuple[str | UUID | TracerSession, str | UUID | TracerSession], /, evaluators: Sequence[Callable[[Sequence[Run], Example | None], ComparisonEvaluationResult | dict | Awaitable[ComparisonEvaluationResult | dict]]], experiment_prefix: str | None = None, description: str | None = None, max_concurrency: int = 5, client: Client | None = None, metadata: dict | None = None, load_nested: bool = False, randomize_order: bool = False) ComparativeExperimentResults [source]#
Evaluate existing experiment runs against each other.
This lets you use pairwise preference scoring to generate more reliable feedback in your experiments.
- Parameters:
experiments (Tuple[Union[str, uuid.UUID], Union[str, uuid.UUID]]) – The identifiers of the experiments to compare.
evaluators (Sequence[COMPARATIVE_EVALUATOR_T]) – A list of evaluators to run on each example.
experiment_prefix (Optional[str]) – A prefix to provide for your experiment name. Defaults to None.
description (Optional[str]) – A free-form text description for the experiment.
max_concurrency (int) – The maximum number of concurrent evaluations to run. Defaults to 5.
client (Optional[langsmith.Client]) – The LangSmith client to use. Defaults to None.
metadata (Optional[dict]) – Metadata to attach to the experiment. Defaults to None.
load_nested (bool) – Whether to load all child runs for the experiment. Default is to only load the top-level root runs.
randomize_order (bool) – Whether to randomize the order of the outputs for each evaluation. Default is False.
- Returns:
The results of the comparative evaluation.
- Return type:
Examples
Suppose you want to compare two prompts to see which one is more effective. You would first prepare your dataset:
>>> from typing import Sequence >>> from langsmith import Client >>> from langsmith.evaluation import evaluate >>> from langsmith.schemas import Example, Run >>> client = Client() >>> dataset = client.clone_public_dataset( ... "https://smith.langchain.com/public/419dcab2-1d66-4b94-8901-0357ead390df/d" ... ) >>> dataset_name = "Evaluate Examples"
Then you would run your different prompts: >>> import functools >>> import openai >>> from langsmith.evaluation import evaluate >>> from langsmith.wrappers import wrap_openai >>> oai_client = openai.Client() >>> wrapped_client = wrap_openai(oai_client) >>> prompt_1 = “You are a helpful assistant.” >>> prompt_2 = “You are an exceedingly helpful assistant.” >>> def predict(inputs: dict, prompt: str) -> dict: … completion = wrapped_client.chat.completions.create( … model=”gpt-3.5-turbo”, … messages=[ … {“role”: “system”, “content”: prompt}, … { … “role”: “user”, … “content”: f”Context: {inputs[‘context’]}” … f”nninputs[‘question’]”, … }, … ], … ) … return {“output”: completion.choices[0].message.content} >>> results_1 = evaluate( … functools.partial(predict, prompt=prompt_1), … data=dataset_name, … description=”Evaluating our basic system prompt.”, … blocking=False, # Run these experiments in parallel … ) # doctest: +ELLIPSIS View the evaluation results for experiment:… >>> results_2 = evaluate( … functools.partial(predict, prompt=prompt_2), … data=dataset_name, … description=”Evaluating our advanced system prompt.”, … blocking=False, … ) # doctest: +ELLIPSIS View the evaluation results for experiment:… >>> results_1.wait() >>> results_2.wait() >>> import time >>> time.sleep(10) # Wait for the traces to be fully processed
Finally, you would compare the two prompts directly:
>>> import json >>> from langsmith.evaluation import evaluate_comparative >>> def score_preferences(runs: list, example: schemas.Example): ... assert len(runs) == 2 # Comparing 2 systems ... assert isinstance(example, schemas.Example) ... assert all(run.reference_example_id == example.id for run in runs) ... pred_a = runs[0].outputs["output"] ... pred_b = runs[1].outputs["output"] ... ground_truth = example.outputs["answer"] ... tools = [ ... { ... "type": "function", ... "function": { ... "name": "rank_preferences", ... "description": "Saves the prefered response ('A' or 'B')", ... "parameters": { ... "type": "object", ... "properties": { ... "reasoning": { ... "type": "string", ... "description": "The reasoning behind the choice.", ... }, ... "preferred_option": { ... "type": "string", ... "enum": ["A", "B"], ... "description": "The preferred option, either 'A' or 'B'", ... }, ... }, ... "required": ["preferred_option"], ... }, ... }, ... } ... ] ... completion = openai.Client().chat.completions.create( ... model="gpt-3.5-turbo", ... messages=[ ... {"role": "system", "content": "Select the better response."}, ... { ... "role": "user", ... "content": f"Option A: {pred_a}" ... f"\n\nOption B: {pred_b}" ... f"\n\nGround Truth: {ground_truth}", ... }, ... ], ... tools=tools, ... tool_choice={ ... "type": "function", ... "function": {"name": "rank_preferences"}, ... }, ... ) ... tool_args = completion.choices[0].message.tool_calls[0].function.arguments ... loaded_args = json.loads(tool_args) ... preference = loaded_args["preferred_option"] ... comment = loaded_args["reasoning"] ... if preference == "A": ... return { ... "key": "ranked_preference", ... "scores": {runs[0].id: 1, runs[1].id: 0}, ... "comment": comment, ... } ... else: ... return { ... "key": "ranked_preference", ... "scores": {runs[0].id: 0, runs[1].id: 1}, ... "comment": comment, ... } >>> def score_length_difference(runs: list, example: schemas.Example): ... # Just return whichever response is longer. ... # Just an example, not actually useful in real life. ... assert len(runs) == 2 # Comparing 2 systems ... assert isinstance(example, schemas.Example) ... assert all(run.reference_example_id == example.id for run in runs) ... pred_a = runs[0].outputs["output"] ... pred_b = runs[1].outputs["output"] ... if len(pred_a) > len(pred_b): ... return { ... "key": "length_difference", ... "scores": {runs[0].id: 1, runs[1].id: 0}, ... } ... else: ... return { ... "key": "length_difference", ... "scores": {runs[0].id: 0, runs[1].id: 1}, ... } >>> results = evaluate_comparative( ... [results_1.experiment_name, results_2.experiment_name], ... evaluators=[score_preferences, score_length_difference], ... client=client, ... ) View the pairwise evaluation results at:... >>> eval_results = list(results) >>> assert len(eval_results) >= 10 >>> assert all( ... "feedback.ranked_preference" in r["evaluation_results"] ... for r in eval_results ... ) >>> assert all( ... "feedback.length_difference" in r["evaluation_results"] ... for r in eval_results ... )