Skip to main content

How to run pairwise evaluations

Key concepts

LangSmith supports evaluating existing experiments in a comparative manner. This allows you to score the outputs from multiple experiments against each other, rather than being confined to evaluating outputs one at a time. Think LMSYS Chatbot Arena - this is the same concept! To do this, use the evaluate_comparative / evaluateComparative function with two existing experiments.

If you haven't already created experiments to compare, check out our quick start or our how-to guide to get started with evaluations.

evaluate_comparative args​

note

Pairwise evaluations require langsmith SDK Python version >=0.1.145 or JS version >=0.1.24.

At its simplest, evaluate_comparative / evaluateComparative function takes the following arguments:

ArgumentDescription
experimentsA list of the two existing experiments you would like to evaluate against each other. These can be uuids or experiment names.
evaluatorsA list of the pairwise evaluators that you would like to attach to this evaluation. See the section below for how to define these.

Along with these, you can also pass in the following optional args:

ArgumentDescription
randomize_order / randomizeOrderAn optional boolean indicating whether the order of the outputs should be randomized for each evaluation. This is a strategy for minimizing positional bias in your prompt: often, the LLM will be biased towards one of the responses based on the order. This should mainly be addressed via prompt engineering, but this is another optional mitigation. Defaults to False.
experiment_prefix / experimentPrefixA prefix to be attached to the beginning of the pairwise experiment name. Defaults to None.
descriptionA description of the pairwise experiment. Defaults to None.
max_concurrency / maxConcurrencyThe maximum number of concurrent evaluations to run. Defaults to 5.
clientThe LangSmith client to use. Defaults to None.
metadataMetadata to attach to your pairwise experiment. Defaults to None.
load_nested / loadNestedWhether to load all child runs for the experiment. When False, only the root trace will be passed to your evaluator. Defaults to False.

Define a pairwise evaluator​

Pairwise evaluators are just functions with an expected signature.

Evaluator args​

Custom evaluator functions must have specific argument names. They can take any subset of the following arguments:

Python and JS/TS

  • runs: list[langsmith.schemas.Run]: A two-item list of the full Run objects generated by the two experiments on the given example. Use this if you need access to intermediate steps or metadata about each run.
  • example: langsmith.schemas.Example: The full dataset Example, including the example inputs, outputs (if available), and metdata (if available).

Currently Python only

  • inputs: dict: A dictionary of the inputs corresponding to a single example in a dataset.
  • outputs: list[dict]: A two-item list of the outputs produced by each experiment on the given inputs.
  • reference_outputs: dict: A dictionary of the reference outputs associated with the example, if available.

For most Python use cases you'll only need inputs, outputs, and reference_outputs. run and example are useful only if you need some extra trace or example metadata outside of the actual inputs and outputs of the application.

Evaluator output​

Custom evaluators are expected to return one of the following types:

Python and JS/TS

  • dict: dictionary with keys:
    • key, which represents the feedback key that will be logged
    • scores, which is a mapping from run ID to score for that run.

Currently Python only

  • list[int | float | bool]: a two-item list of scores. The list is assumed to have the same order as the runs / outputs evaluator args. The evaluator function name is used for the feedback key.

Note that you should choose a feedback key that is distinct from standard feedbacks on your run. We recommend prefixing pairwise feedback keys with pairwise_ or ranked_.

Run a pairwise evaluation​

The following example uses a prompt which asks the LLM to decide which is better between two AI assistant responses. It uses structured output to parse the AI's response: 0, 1, or 2.

Optional LangChain Usage

In the Python example below, we are pulling this structured prompt from the LangChain Hub and using it with a LangChain chat model wrapper.

Usage of LangChain is totally optional. To illustrate this point, the TypeScript example uses the OpenAI SDK directly.

Requires langsmith>=0.1.145

from langchain import hub
from langchain.chat_models import init_chat_model
from langsmith import evaluate_comparative

prompt = hub.pull("langchain-ai/pairwise-evaluation-2")
model = init_chat_model("gpt-4o")
chain = prompt | model

def ranked_preference(inputs: dict, outputs: list[dict]) -> list:
    response = chain.invoke({
        "question": inputs["question"],
        "answer_a": outputs[0].get("output", "N/A"),
        "answer_b": outputs[1].get("output", "N/A"),
    })
    preference = response["Preference"]

    if preference == 1:
        scores = [1, 0]
    elif preference == 2:
        scores = [0, 1]
    else:
        scores = [0, 0]
    return scores

evaluate_comparative(
    # Replace the following array with the names or IDs of your experiments
    ["my-experiment-name-1", "my-experiment-name-2"],
    evaluators=[ranked_preference],
)

View pairwise experiments​

Navigate to the "Pairwise Experiments" tab from the dataset page:

Pairwise Experiments Tab

Click on a pairwise experiment that you would like to inspect, and you will be brought to the Comparison View:

Pairwise Comparison View

You may filter to runs where the first experiment was better or vice versa by clicking the thumbs up/thumbs down buttons in the table header:

Pairwise Filtering

Was this page helpful?


You can leave detailed feedback on GitHub.