Register Custom Metrics

​

Registered Scorers

We support registering a scorer such that it can be reused across various runs, projects, modules, and users within your organization. Registered Scorers are run in the backend in an isolated environment that has access to a predefined set of libraries and packages.

​

Creating Your Registered Scorer

To define a registered scorer, create a Python file that has at least 2 functions and follow the function signatures as described below:

1. scorer_fn: The scorer function is provided the row-wise inputs and is expected to generate outputs for each response. The expected signature for this function is:

 def scorer_fn(*, index: Union[int, str], node_input: str, node_output: str, **kwargs: Any) -> Union[float, int, bool, str, None]:
    ...

We support output of a floating points, integers, boolean values, and strings. Your scorer_fn must accept **kwargs as the last parameter so that your registered scorer is forward-compatible.

Here is an example with the full list of parameters supported currently. This example checks the output vs the ground truth and returns the absolute difference in length:

 def scorer_fn(*, index: Union[int, str], node_input: str, node_output: str, node_name: Optional[str], node_type: Optional[str], node_id: Optional[UUID], dataset_variables: Dict[str, str], **kwargs: Any) -> Union[float, int, bool, str, None]:
    ground_truth = dataset_variables.get("target", "") # ground truth for row if it was provided.
    return abs(len(node_output) - len(ground_truth))

node_name, node_type, and node_id are all specific to workflows/multi step chains. dataset_variables contains key-value pairs of variables that are passed in from the dataset in prompt evaluation runs, but can also be used to get the target/ground truth in multi step runs. Dataset variables are not available for Evaluate workflows / Observe.

The index parameter is the index of the row in the dataset, node_input is the input to the node, and node_output is the output from the node.

2. aggregator_fn: The aggregator function is only used in Evaluate, not Observe. The aggregator function takes in an array of the row-wise outputs from your scorer and allows you to generate aggregates from those. The expected signature for the aggregator function is:

   
       def aggregator_fn(*, scores: List[Union[float, int, bool, str, None]]) -> Dict[str, Union[float, int, bool, str, None]]:
           ...
   

   For aggregated values that you want to output from your scorer, return them as key-value pairs with the key corresponding to the label and the value.

3. (Optional, but recommended) score_type: The scorer_type function is used to define the Type of the score that your scorer generates. The expected signature for this function is:

   
       def score_type() -> Type[float] | Type[int] | Type[str] | Type[bool]:
       ...
   

   Note that the return type is a Type object like float, not the actual type itself. Defining this function is necessary for sorting and filtering by scores to work correctly. If you don’t define this function, the scorer is assumed to generate float scores by default.

4. (Optional) scoreable_node_types_fn: If you want to restrict your scorer to only run on specific node types, you can define this function which returns a list of node types that your scorer should run on. The expected signature for this function is:

   def scoreable_node_types_fn() -> List[str]:
           ...
   

   If you don’t define this function, your scorer will run on llm and chat nodes by default.

   Here’s an example of a scoreable_node_types_fn that restricts the scorer to only run on retriever nodes:

   def scoreable_node_types_fn() -> List[str]:
       return ["retriever"]
   

5. (Optional) include_llm_credentials: If you want access to the LLM credentials for the user who created the Observe project / Evaluate run during the execution of the registered scorer. This is expected to be set as a boolean value. OpenAI credentials are the only ones that are currently supported. By default, it is assumed to be False. The expected signature for this property is:

   include_llm_credentials = True
   

   If you don’t define this function, your scorer will not have access to the LLM credentials by default. If you do enable it, the credentials will be included in calls to scorer_fn at the keyword argument credentials. The credentials will be a dictionary with the keys as the name of the integration, if available, and values as the credentials. For example, if the user has an OpenAI integration, the credentials will be:

   {
     "openai": {
       "api_key": "foo", // str
       "organization": "my-org-id" // Optional[str]
     }
   }
   

​

Registering Your Scorer

Once you’ve created your scorer file, you can register it with the name and the scorer file:

    registered_scorer = pq.register_scorer(scorer_name="my-scorer", scorer_file="/path/to/scorer/file.py")

The name you choose here will be the name with which the values for this scorer appear in the UI later.

​

Using Your Registered Scorer

To use your scorer during a prompt run (or sweep), simply pass it in alongside any of the other scorers:

    pq.run(..., scorers=[registered_scorer])

If you created your registered scorer in a previous session, you can also just pass in the name to the scorer instead of the object as:

    pq.run(..., scorers=["my-scorer"])

​

Example

For example, let’s say we wanted to create a custom metric that measured the length of the response. In our Python environment, we would define a scorer_fn function, and an aggregator_fn function.

1. Create a scorer.py file:

    from typing import List, Dict, Type


    def scorer_fn(*, response: str, **kwargs) -> int:
        return len(response)


    def aggregator_fn(*, scores: List[str]) -> Dict[str, int]:
        return {
            "Total Response Length": sum(scores),
            "Average Response Length": sum(scores) / len(scores),
        }

    def score_type() -> Type:
        return int

    def scoreable_node_types_fn() -> List[str]:
        return ["llm", "chat"]

1. Register the scorer:

       pq.register_scorer("response_length", "scorer.py")
   

2. Use the scorer in your prompt run:

       pq.run(..., scorers=["response_length"])
   

​

Execution Environment

Your scorer will be executed in a Python 3.10 environment. The Python libraries available for your use are:

    numpy~=1.26.4
    pandas~=2.2.2
    pydantic~=2.7.1
    scikit-learn~=1.4.2
    tensorflow~=2.16.1
    networkx
    openai

Please note that we regularly update the minor and patch versions of these packages. Major version updates are infrequent but if a library is critical to your scorer, please let us know and we’ll provide 1+ week of warning before updating the major versions for those.

​

What if I need to use other libraries or packages?

If you need to use other libraries or packages, you may use ‘Custom Scorers’. Custom Scorers are run on your notebook environment. Because they run locally, they won’t be available for runs created from the UI or for Observe projects.

​

How do I create a local “Custom Scorer”?

Custom scorers can be created from two Python functions (executor and aggrator function as defined below). Common types include:

1. Heuristics/custom rules: checking for regex matches or presence/absence of certain keywords or phrases.

2. model-guided: utilizing a pre-trained model to check for specific entities (e.g. PERSON, ORG), or asking an LLM to grade the quality of the output.

For example, for that registered scorer we created to calculate response length, here is the custom scorer equivalent:

Note that the naming of the functions are different: they are **executor** and **aggregator** instead of scorer_fn and aggregator_fn.

    from typing import Dict, List
    from promptquality import PromptRow

    def executor(row: PromptRow) -> float:
      return len(row.response)

    def aggregator_fn(scores: float, indices: List[int]) -> Dict[str, float]:
      return {'Total Response Length': sum(scores),
              # You can have multiple aggregate summaries for your metric.
              'Average Response Length': sum(scores)/len(scores)}

    my_scorer = pq.CustomScorer(name='Response Length', executor=executor, aggregator=aggregator_fn)

To use your scorer, you can just pass it through your scorers parameter inside pq.run or pq.run_sweep, pq.EvaluateRun, or pq.GalileoPromptCallback:

    template = "Explain {topic} to me like I'm a 5 year old"

    data = {"topic": ["Quantum Physics", "Politics", "Large Language Models"]}

    pq.run(template = template, dataset = data, scorers=[my_scorer])

Note that custom scorer can only be used in the Evaluate module - if you want to use a custom metric to evaluate live traffic (Observe module), you’ll need to use the registered scorers below.