Much like tests in traditional software, evals are an important part of bringing LLM applications to production. The goal of this package is to help provide a starting point for you to write evals for your LLM applications, from which you can write more c
npm install openevalsMuch like tests in traditional software, evals are an important part of bringing LLM applications to production.
The goal of this package is to help provide a starting point for you to write evals for your LLM applications, from which
you can write more custom evals specific to your application.
If you are looking for evals specific to evaluating LLM agents, please check out agentevals.
> [!TIP]
> If you'd like to follow along with a video walkthrough, click the image below:
> 
To get started, install openevals:
``bash`
npm install openevals @langchain/core
This quickstart will use an evaluator powered by OpenAI's o3-mini model to judge your results, so you'll need to set your OpenAI API key as an environment variable:
`bash`
export OPENAI_API_KEY="your_openai_api_key"
Once you've done this, you can run your first eval:
`ts
import { createLLMAsJudge, CONCISENESS_PROMPT } from "openevals";
const concisenessEvaluator = createLLMAsJudge({
// CONCISENESS_PROMPT is just an f-string
prompt: CONCISENESS_PROMPT,
model: "openai:o3-mini",
});
const inputs = "How is the weather in San Francisco?"
// These are fake outputs, in reality you would run your LLM-based system to get real outputs
const outputs = "Thanks for asking! The current weather in San Francisco is sunny and 90 degrees."
// When calling an LLM-as-judge evaluator, parameters are formatted directly into the prompt
const evalResult = await concisenessEvaluator({
inputs,
outputs,
});
console.log(evalResult);
`
`
{
key: 'score',
score: false,
comment: 'The output includes an unnecessary greeting ("Thanks for asking!") and extra..'
}
This is an example of a reference-free evaluator - some other evaluators may accept slightly different parameters such as a required reference output. LLM-as-judge evaluators will attempt to format any passed parameters into their passed prompt, allowing you to flexibly customize criteria or add other fields.
See the LLM-as-judge section for more information on how to customize the scoring to output float values rather than just True/False, the model, or the prompt!
- Installation
- Evaluators
-
- Prebuilt prompts
- Correctness
- Conciseness
- Hallucination
- Toxicity
- Answer relevance
- Plan adherence
- Customizing prompts
- Customizing the model
- Customizing output score values
- Customizing output schema
-
- Evaluating structured output with exact match
- Evaluating structured output with LLM-as-a-Judge
-
- Correctness
- Helpfulness
- Groundedness
- Retrieval relevance
- Retrieval relevance with LLM as judge
- Retrieval relevance with string evaluators
-
- Extracting code outputs
- Pyright (Python-only)
- Mypy (Python-only)
- TypeScript type-checking (TypeScript-only)
- LLM-as-judge for code
-
- Sandboxed Pyright (Python-only)
- Sandboxed TypeScript type-checking (TypeScript-only)
- Sandboxed Execution
-
- Exact Match
- Levenshtein Distance
- Embedding Similarity
- Agent evals
- Creating your own
- Python Async Support
- Multiturn Simulation
- Simulating users
- Multiturn simulation with LangGraph
- LangSmith Integration
- Pytest or Vitest/Jest
- Evaluate
You can install openevals like this:
`bash`
npm install openevals @langchain/core
For LLM-as-judge evaluators, you will also need an LLM client. By default, openevals will use LangChain chat model integrations and comes with langchain_openai installed by default. However, if you prefer, you may use the OpenAI client directly:
`bash`
npm install openai
It is also helpful to be familiar with some evaluation concepts.
One common way to evaluate an LLM app's outputs is to use another LLM as a judge. This is generally a good starting point for evals.
This package contains the create_llm_as_judge function, which takes a prompt and a model as input, and returns an evaluator function
that handles converting parameters into strings and parsing the judge LLM's outputs as a score.
To use the create_llm_as_judge function, you need to provide a prompt and a model. To get started, OpenEvals has some prebuilt prompts in the openevals.prompts module that you can use out of the box. Here's an example:
`ts
import { createLLMAsJudge, CORRECTNESS_PROMPT } from "openevals";
const correctnessEvaluator = createLLMAsJudge({
prompt: CORRECTNESS_PROMPT,
model: "openai:o3-mini",
});
`
Note that CORRECTNESS_PROMPT is a simple f-string that you can log and edit as needed for your specific use case:
`ts`
console.log(CORRECTNESS_PROMPT);
`
You are an expert data labeler evaluating model outputs for correctness. Your task is to assign a score based on the following rubric:
A correct answer:
- Provides accurate and complete information
...
{inputs}
...
`
By convention, we generally suggest sticking to inputs, outputs, and reference_outputs as the names of the parameters for LLM-as-judge evaluators, but these will be directly formatted into the prompt so you can use any variable names you want.
openevals includes a prebuilt prompt for create_llm_as_judge that scores the conciseness of an LLM's output. It takes inputs and outputs as parameters.
`ts
import { createLLMAsJudge, CONCISENESS_PROMPT } from "openevals";
const concisenessEvaluator = createLLMAsJudge({
prompt: CONCISENESS_PROMPT,
feedbackKey: "conciseness",
model: "openai:o3-mini",
});
const inputs = "How is the weather in San Francisco?"
const outputs = "Thanks for asking! The current weather in San Francisco is sunny and 90 degrees."
const evalResult = await concisenessEvaluator({
inputs,
outputs,
});
console.log(evalResult);
`
`
{
key: 'conciseness',
score: false,
comment: '...'
}
openevals includes a prebuilt prompt for create_llm_as_judge that scores the correctness of an LLM's output. It takes inputs, outputs, and optionally, reference_outputs as parameters.
`ts
import { createLLMAsJudge, CORRECTNESS_PROMPT } from "openevals";
const correctnessEvaluator = createLLMAsJudge({
prompt: CORRECTNESS_PROMPT,
feedbackKey: "correctness",
model: "openai:o3-mini",
});
const inputs = "How much has the price of doodads changed in the past year?";
const outputs = "Doodads have increased in price by 10% in the past year.";
const referenceOutputs = "The price of doodads has decreased by 50% in the past year.";
const evalResult = await correctnessEvaluator({
inputs,
outputs,
referenceOutputs,
});
console.log(evalResult);
`
`
{
key: 'correctness',
score: false,
comment: '...'
}
openevals includes a prebuilt prompt for create_llm_as_judge that scores the hallucination of an LLM's output. It takes inputs, outputs, and optionally, context as parameters.
`ts
import { createLLMAsJudge, HALLUCINATION_PROMPT } from "openevals";
const hallucinationEvaluator = createLLMAsJudge({
prompt: HALLUCINATION_PROMPT,
feedbackKey: "hallucination",
model: "openai:o3-mini",
});
const inputs = "What is a doodad?"
const outputs = "I know the answer. A doodad is a kitten."
const context = "A doodad is a self-replicating swarm of nanobots. They are extremely dangerous and should be avoided at all costs. Some safety precautions when working with them include wearing gloves and a mask."
const evalResult = await hallucinationEvaluator({
inputs,
outputs,
context,
referenceOutputs: "",
});
console.log(evalResult);
`
`
{
key: 'hallucination',
score: false,
comment: '...'
}
openevals includes a prebuilt prompt for create_llm_as_judge that scores the toxicity of an LLM's output. It takes inputs and outputs as parameters.
`ts
import { createLLMAsJudge, TOXICITY_PROMPT } from "openevals";
const toxicityEvaluator = createLLMAsJudge({
prompt: TOXICITY_PROMPT,
feedbackKey: "toxicity",
model: "openai:o3-mini",
});
const inputs = "What is a doodad?";
const outputs = "You stink!";
const evalResult = await toxicityEvaluator({
inputs,
outputs,
referenceOutputs: "",
});
console.log(evalResult);
`
`
{
key: 'toxicity',
score: true,
comment: '...'
}
openevals includes a prebuilt prompt for create_llm_as_judge that scores how relevant an LLM's output is to the input. It takes inputs and outputs as parameters.
`ts
import { createLLMAsJudge, ANSWER_RELEVANCE_PROMPT } from "openevals";
const answerRelevanceEvaluator = createLLMAsJudge({
prompt: ANSWER_RELEVANCE_PROMPT,
feedbackKey: "answer_relevance",
model: "openai:o3-mini",
});
const inputs = "What is a doodad?";
const outputs = "A doodad is a thingy.";
const evalResult = await answerRelevanceEvaluator({
inputs,
outputs,
referenceOutputs: "",
});
console.log(evalResult);
`
`
{
key: 'answer_relevance',
score: true,
comment: '...'
}
openevals includes a prebuilt prompt for create_llm_as_judge that scores how well an LLM's output adheres to a given plan. It takes inputs, outputs, and a plan as parameters.
For this prompt, your outputs should also include a set of steps that the LLM-as-judge can compare against the passed plan.
`ts
import { createLLMAsJudge, PLAN_ADHERENCE_PROMPT } from "openevals";
const planAdherenceEvaluator = createLLMAsJudge({
prompt: PLAN_ADHERENCE_PROMPT,
feedbackKey: "plan_adherence",
model: "openai:o3-mini",
});
const inputs = "What is a doodad?";
const plan = {
steps: [
"Research the definition of a doodad",
"Provide a concise answer",
]
};
const outputs = {
steps: [
"I've done my research on Google and have concluded that a doodad is a thingy."
],
answer: "A doodad is a thingy."
};
const evalResult = await planAdherenceEvaluator({
inputs,
outputs,
referenceOutputs: "",
plan,
});
console.log(evalResult);
`
`
{
key: 'plan_adherence',
score: true,
comment: '...'
}
The prompt parameter for create_llm_as_judge may be an f-string, LangChain prompt template, or a function that takes kwargs and returns a list of formatted messages.
Though we suggest sticking to conventional names (inputs, outputs, and reference_outputs) as prompt variables, your prompts can also require additional variables. You would then pass these extra variables when calling your evaluator function. Here's an example of a prompt that requires an extra variable named context:
`ts
import { createLLMAsJudge } from "openevals";
const MY_CUSTOM_PROMPT =
Use the following context to help you evaluate for hallucinations in the output:
{context}
{inputs}
;
const customPromptEvaluator = createLLMAsJudge({
prompt: MY_CUSTOM_PROMPT,
model: "openai:o3-mini",
});
const inputs = "What color is the sky?"
const outputs = "The sky is red."
const evalResult = await customPromptEvaluator({
inputs,
outputs,
});
`
The following options are also available for string prompts:
- system: a string that sets a system prompt for the judge model by adding a system message before other parts of the prompt.few_shot_examples
- : a list of example dicts that are appended to the end of the prompt. This is useful for providing the judge model with examples of good and bad outputs. The required structure looks like this:
`ts`
const fewShotExamples = [
{
inputs: "What color is the sky?",
outputs: "The sky is red.",
reasoning: "The sky is red because it is early evening.",
score: 1,
}
]
These will be appended to the end of the final user message in the prompt.
#### Customizing with LangChain prompt templates
You can also pass a LangChain prompt template if you want more control over formatting. Here's an example that uses mustache formatting instead of f-strings:
`ts
import { createLLMAsJudge } from "openevals";
import { ChatPromptTemplate } from "@langchain/core/prompts";
const inputs = { a: 1, b: 2 };
const outputs = { a: 1, b: 2 };
const prompt = ChatPromptTemplate.fromMessages([
["system", "You are an expert at determining if two objects are equal."],
["user", "Are these two equal? {{inputs}} {{outputs}}"],
], { templateFormat: "mustache" });
const evaluator = createLLMAsJudge({
prompt,
model: "openai:o3-mini",
feedbackKey: "equality",
});
const result = await evaluator({ inputs, outputs });
`
`
{
key: 'equality',
score: true,
comment: '...'
}
You can also pass in a function that takes your LLM-as-judge inputs as kwargs and returns formatted chat messages.
There are a few ways you can customize the model used for evaluation. You can pass a string formatted as PROVIDER:MODEL (e.g. model=anthropic:claude-3-5-sonnet-latest) as the model, in which case the package will attempt to import and initialize a LangChain chat model instance. This requires you to install the appropriate LangChain integration package installed. Here's an example:
`bash`
npm install @langchain/anthropic
`ts
import { createLLMAsJudge, CORRECTNESS_PROMPT } from "openevals";
const anthropicEvaluator = createLLMAsJudge({
prompt: CORRECTNESS_PROMPT,
model: "anthropic:claude-3-5-sonnet-latest",
});
`
You can also directly pass a LangChain chat model instance as judge. Note that your chosen model must support structured output:
`ts
import { createLLMAsJudge, CORRECTNESS_PROMPT } from "openevals";
import { ChatAnthropic } from "@langchain/anthropic";
const anthropicEvaluator = createLLMAsJudge({
prompt: CORRECTNESS_PROMPT,
judge: new ChatAnthropic({ model: "claude-3-5-sonnet-latest", temperature: 0.5 }),
});
`
This is useful in scenarios where you need to initialize your model with specific parameters, such as temperature or alternate URLs if using models through a service like Azure.
Finally, you can pass a model name as model and a judge parameter set to an OpenAI client instance:
`bash`
npm install openai
`ts
import { OpenAI } from "openai";
import { createLLMAsJudge, CORRECTNESS_PROMPT } from "openevals";
const openaiEvaluator = createLLMAsJudge({
prompt: CORRECTNESS_PROMPT,
model: "gpt-4o-mini",
judge: new OpenAI(),
});
`
There are two fields you can set to customize the outputted scores of your evaluator:
- continuous: a boolean that sets whether the evaluator should return a float score somewhere between 0 and 1 instead of a binary score. Defaults to False.choices
- : a list of floats that sets the possible scores for the evaluator.
These parameters are mutually exclusive. When using either of them, you should make sure that your prompt is grounded in information on what specific scores mean - the prebuilt ones in this repo do not have this information!
For example, here's an example of how to define a less harsh definition of correctness that only penalizes incorrect answers by 50% if they are on-topic:
`ts
import { createLLMAsJudge } from "openevals";
const MY_CUSTOM_PROMPT =
You are an expert data labeler evaluating model outputs for correctness. Your task is to assign a score based on the following rubric:
Assign a score of 0, .5, or 1 based on the following criteria:
- 0: The answer is incorrect and does not mention doodads
- 0.5: The answer mentions doodads but is otherwise incorrect
- 1: The answer is correct and mentions doodads
{inputs}
{reference_outputs};
const customEvaluator = createLLMAsJudge({
prompt: MY_CUSTOM_PROMPT,
choices: [0.0, 0.5, 1.0],
model: "openai:o3-mini",
});
const result = await customEvaluator({
inputs: "What is the current price of doodads?",
outputs: "The price of doodads is $10.",
reference_outputs: "The price of doodads is $15.",
});
console.log(result);
`
`
{
'key': 'score',
'score': 0.5,
'comment': 'The provided answer mentioned doodads but was incorrect.'
}
Finally, if you would like to disable justifications for a given score, you can set use_reasoning=False when creating your evaluator.
If you need to change the structure of the raw output generated by the LLM, you can also pass a custom output schema into your LLM-as-judge evaluator as output_schema (Python) / outputSchema (TypeScript). This may be helpful for specific prompting strategies or if you would like to extract multiple metrics at the same time rather than over multiple calls.
> [!CAUTION]
> Passing output_schema changes the return value of the evaluator to match the passed output_schema value instead of the typical OpenEvals format.
> We recommend sticking with the default schema if you do not specifically need additional properties.
For Python, output_schema may be:
- A TypedDict instance
- A Pydantic model
- JSON schema
- OpenAI's structured output format
For TypeScript, outputSchema may be:
- A Zod object
- JSON schema
- OpenAI's structured output format
Note that if you are using an OpenAI client directly, only JSON schema and OpenAI's structured output format.
Here's an example:
`ts
import { z } from "zod";
import { createLLMAsJudge } from "openevals";
const equalitySchema = z.object({
equality_justification: z.string(),
are_equal: z.boolean(),
})
const inputs = "The rain in Spain falls mainly on the plain.";
const outputs = "The rain in Spain falls mainly on the plain.";
const llmAsJudge = createLLMAsJudge({
prompt: "Are the following two values equal? {inputs} {outputs}",
model: "openai:o3-mini",
outputSchema: equalitySchema,
});
const evalResult = await llmAsJudge({ inputs, outputs });
console.log(evalResult);
`
`
{
'equality_justification': 'The values are equal because they have the same properties with identical values.',
'are_equal': True,
}
#### Logging feedback with custom output schemas
If you are using an OpenEvals evaluator with LangSmith's pytest or Vitest/Jest runners, you will need to manually log feedback keys.
If you are using evaluate, you will need to wrap your evaluator in another function that maps your evaluator return value to feedback in the right format.
#### Structured prompts
Passing in a pulled prompt from the LangChain prompt hub that has an output schema set will also change the output schema for the LLM-as-judge evaluator.
Two very common use cases for LLMs are extracting structured output from documents and tool calling. Both of these require the LLM
to respond in a structured format. This package provides a prebuilt evaluator to help you evaluate these use cases, and is flexible
to work for a variety of extraction/tool calling use cases.
You can use the create_json_match_evaluator evaluator in two ways:
1. To perform an exact match of the outputs to reference outputs
2. Using LLM-as-a-judge to evaluate the outputs based on a provided rubric.
Note that this evaluator may return multiple scores based on key and aggregation strategy, so the result will be an array of scores rather than a single one.
Use exact match evaluation when there is a clear right or wrong answer. A common scenario is text extraction from images or PDFs where you expect specific values.
`ts
import { createJsonMatchEvaluator } from "openevals";
import { OpenAI } from "openai";
const outputs = [
{a: "Mango, Bananas", b: 2},
{a: "Apples", b: 2, c: [1,2,3]},
]
const reference_outputs = [
{a: "Mango, Bananas", b: 2},
{a: "Apples", b: 2, c: [1,2,4]},
]
const client = new OpenAI();
const evaluator = createJsonMatchEvaluator({
// How to aggregate feedback keys in each element of the list: "average", "all", or None
// "average" returns the average score. "all" returns 1 only if all keys score 1; otherwise, it returns 0. None returns individual feedback chips for each key
aggregator="all",
// Remove if evaluating a single structured output. This aggregates the feedback keys across elements of the list. Can be "average" or "all". Defaults to "all". "all" returns 1 if each element of the list is 1; if any score is not 1, it returns 0. "average" returns the average of the scores from each element.
list_aggregator="average",
// The keys to ignore during evaluation. Any key not passed here or in rubric will be evaluated using an exact match comparison to the reference outputs
exclude_keys=["a"],
// The provider and name of the model to use
judge: client,
model: "openai:o3-mini",
})
// Invoke the evaluator with the outputs and reference outputs
const result = await evaluator({
outputs,
reference_outputs,
})
console.log(result)
`
For the first element, "b" will be 1 and the aggregator will return a score of 1
For the second element, "b" will be 1, "c" will be 0 and the aggregator will return a score of 0
Therefore, the list aggregator will return a final score of 0.5.
`
[
{
'key': 'json_match:all',
'score': 0.5,
'comment': None,
}
]
Use LLM-as-a-judge to evaluate structured output or tools calls when the criteria is more subjective (for example the output is a kind of fruit or mentions all the fruits).
`ts
import { createJsonMatchEvaluator } from "openevals";
import { OpenAI } from "openai";
const outputs = [
{a: "Mango, Bananas", b: 2},
{a: "Apples", b: 2, c: [1,2,3]},
]
const reference_outputs = [
{a: "Bananas, Mango", b: 2},
{a: "Apples, Strawberries", b: 2},
]
const client = new OpenAI();
const evaluator = createJsonMatchEvaluator({
// How to aggregate feedback keys in each element of the list: "average", "all", or None
// "average" returns the average score. "all" returns 1 only if all keys score 1; otherwise, it returns 0. None returns individual feedback chips for each key
aggregator="average",
// Remove if evaluating a single structured output. This aggregates the feedback keys across elements of the list. Can be "average" or "all". Defaults to "all". "all" returns 1 if each element of the list is 1; if any score is not 1, it returns 0. "average" returns the average of the scores from each element.
list_aggregator="all",
// The criteria for the LLM judge to use for each key you want evaluated by the LLM
rubric={
a: "Does the answer mention all the fruits in the reference answer?"
},
// The keys to ignore during evaluation. Any key not passed here or in rubric will be evaluated using an exact match comparison to the reference outputsrubric
exclude_keys=["c"],
// The provider and name of the model to use
judge: client,
model: "openai:o3-mini",
// Whether to use reasoning to reason about the keys in . Defaults to True
useReasoning: true
})
// Invoke the evaluator with the outputs and reference outputs
const result = await evaluator({
outputs,
reference_outputs,
})
console.log(result)
`
For the first element, "a" will be 1 since both Mango and Bananas are in the reference output, "b" will be 1 and "d" will be 0. The aggregator will return an average score of 0.6.
For the second element, "a" will be 0 since the reference output doesn't mention all the fruits in the output, "b" will be 1. The aggregator will return a score of 0.5.
Therefore, the list aggregator will return a final score of 0.
`
{
'key': 'json_match:a',
'score': 0,
'comment': None
}
RAG applications in their most basic form consist of 2 steps. In the retrieval step, context is retrieved (often from something like a vector database that a user has prepared ahead of time, though web retrieval use-cases are gaining in popularity as well) to provide the LLM with the information it needs to respond to the user. In the generation step, the LLM uses the retrieved context to formulate an answer.
OpenEvals provides prebuilt prompts and other methods for the following:
1. Correctness
- Evaluates: Final output vs. input + reference answer
- Goal: Measure "how similar/correct is the generated answer relative to a ground-truth answer"
- Requires reference: Yes
2. Helpfulness
- Evaluates: Final output vs. input
- Goal: Measure "how well does the generated response address the initial user input"
- Requires reference: No, because it will compare the answer to the input question
3. Groundedness
- Evaluates: Final output vs. retrieved context
- Goal: Measure "to what extent does the generated response agree with the retrieved context"
- Requires reference: No, because it will compare the answer to the retrieved context
4. Retrieval relevance
- Evaluates: Retrieved context vs. input
- Goal: Measure "how relevant are my retrieved results for this query"
- Requires reference: No, because it will compare the question to the retrieved context
correctness measures how similar/correct a generated answer is to a ground-truth answer. By definition, this requires you to have a reference output to compare against the generated one. It is useful to test your RAG app end-to-end, and does directly take into account context retrieved as an intermediate step.
You can evaluate the correctness of a RAG app's outputs using the LLM-as-judge evaluator alongside the general prebuilt prompt covered above. Here's an example:
`ts
import { createLLMAsJudge, CORRECTNESS_PROMPT } from "openevals";
const correctnessEvaluator = createLLMAsJudge({
prompt: CORRECTNESS_PROMPT,
feedbackKey: "correctness",
model: "openai:o3-mini",
});
const inputs = "How much has the price of doodads changed in the past year?";
const outputs = "Doodads have increased in price by 10% in the past year.";
const referenceOutputs = "The price of doodads has decreased by 50% in the past year.";
const evalResult = await correctnessEvaluator({
inputs,
outputs,
referenceOutputs,
});
console.log(evalResult);
`
`
{
key: 'correctness',
score: false,
comment: '...'
}
For more information on customizing LLM-as-judge evaluators, see these sections.
helpfulness measures how well the generated response addresses the initial user input. It compares the final generated output against the input, and does not require a reference. It's useful to validate that the generation step of your RAG app actually answers the original question as stated, but does not measure that the answer is supported by any retrieved context!
You can evaluate the helpfulness of a RAG app's outputs using the LLM-as-judge evaluator with a prompt like the built-in RAG_HELPFULNESS_PROMPT. Here's an example:
`ts
import { createLLMAsJudge, RAG_HELPFULNESS_PROMPT } from "openevals";
const inputs = {
"question": "Where was the first president of FoobarLand born?",
};
const outputs = {
"answer": "The first president of FoobarLand was Bagatur Askaryan.",
};
const helpfulnessEvaluator = createLLMAsJudge({
prompt: RAG_HELPFULNESS_PROMPT,
feedbackKey: "helpfulness",
model: "openai:o3-mini",
});
const evalResult = await helpfulnessEvaluator({
inputs,
outputs,
});
console.log(evalResult);
`
`
{
'key': 'helpfulness',
'score': False,
'comment': "The question asks for the birthplace of the first president of FoobarLand, but the retrieved outputs only identify the first president as Bagatur and provide an unrelated biographical detail (being a fan of PR reviews). Although the first output is somewhat relevant by identifying the president's name, neither document provides any information about his birthplace. Thus, the outputs do not contain useful information to answer the input question. Thus, the score should be: false."
}
groundedness measures the extent that the generated response agrees with the retrieved context. It compares the final generated output against context fetched during the retrieval step, and verifies that the generation step is properly using retrieved context vs. hallucinating a response or overusing facts from the LLM's base knowledge.
You can evaluate the groundedness of a RAG app's outputs using the LLM-as-judge evaluator with a prompt like the built-in RAG_GROUNDEDNESS_PROMPT. Note that this prompt does not take the example's original inputs into account, only the outputs and their relation to the retrieved context. Thus, unlike some of the other prebuilt prompts, it takes context and outputs as prompt variables:
`ts
import { createLLMAsJudge, RAG_GROUNDEDNESS_PROMPT } from "openevals";
const groundednessEvaluator = createLLMAsJudge({
prompt: RAG_GROUNDEDNESS_PROMPT,
feedbackKey: "groundedness",
model: "openai:o3-mini",
});
const context = {
documents: [
"FoobarLand is a new country located on the dark side of the moon",
"Space dolphins are native to FoobarLand",
"FoobarLand is a constitutional democracy whose first president was Bagatur Askaryan",
"The current weather in FoobarLand is 80 degrees and clear."
],
};
const outputs = {
answer: "The first president of FoobarLand was Bagatur Askaryan.",
};
const evalResult = await groundednessEvaluator({
context,
outputs,
});
console.log(evalResult);
`
`
{
'key': 'groundedness',
'score': true,
'comment': 'The output states, "The first president of FoobarLand was Bagatur Askaryan," which is directly supported by the retrieved context (document 3 explicitly states this fact). There is no addition or modification, and the claim aligns perfectly with the context provided. Thus, the score should be: true.',
'metadata': None
}
retrieval_relevance measures how relevant retrieved context is to an input query. This type of evaluator directly measures the quality of the retrieval step of your app vs. its generation step.
#### Retrieval relevance with LLM-as-judge
You can evaluate the retrieval relevance of a RAG app using the LLM-as-judge evaluator with a prompt like the built-in RAG_RETRIEVAL_RELEVANCE_PROMPT. Note that this prompt does not consider at your actual app's final output, only inputs and the retrieved context. Thus, unlike some of the other prebuilt prompts, it takes context and inputs as prompt variables:
`ts
import { createLLMAsJudge, RAG_RETRIEVAL_RELEVANCE_PROMPT } from "openevals";
const retrievalRelevanceEvaluator = createLLMAsJudge({
prompt: RAG_RETRIEVAL_RELEVANCE_PROMPT,
feedbackKey: "retrieval_relevance",
model: "openai:o3-mini",
});
const inputs = {
question: "Where was the first president of FoobarLand born?",
}
const context = {
documents: [
"FoobarLand is a new country located on the dark side of the moon",
"Space dolphins are native to FoobarLand",
"FoobarLand is a constitutional democracy whose first president was Bagatur Askaryan",
"The current weather in FoobarLand is 80 degrees and clear.",
],
}
const retrievalRelevanceEvaluator = await retrievalRelevanceEvaluator({
inputs,
context,
});
console.log(evalResult);
`
`
{
'key': 'retrieval_relevance',
'score': False,
'comment': "The retrieved context provides some details about FoobarLand – for instance, that it is a new country located on the dark side of the moon and that its first president is Bagatur Askaryan. However, none of the documents specify where the first president was born. Notably, while there is background information about FoobarLand's location, the crucial information about the birth location of the first president is missing. Thus, the retrieved context does not fully address the question. Thus, the score should be: false.",
'metadata': None
}
#### Retrieval relevance with string evaluators
You can also use string evaluators like embedding similarity to measure retrieval relevance without using an LLM. In this case, you should convert your retrieved documents into a string and pass it into your evaluator as outputs, while the original input query will be passed as reference_outputs. The output score and your acceptable threshold will depend on the specific embeddings model you use.
Here's an example:
`ts
import { createEmbeddingSimilarityEvaluator } from "openevals";
import { OpenAIEmbeddings } from "@langchain/openai";
const evaluator = createEmbeddingSimilarityEvaluator({
embeddings: new OpenAIEmbeddings({ model: "text-embedding-3-small" }),
});
const inputs = "Where was the first president of FoobarLand born?";
const context = [
"BazQuxLand is a new country located on the dark side of the moon",
"Space dolphins are native to BazQuxLand",
"BazQuxLand is a constitutional democracy whose first president was Bagatur Askaryan",
"The current weather in BazQuxLand is 80 degrees and clear.",
].join("\n");
const result = await evaluator(
outputs: context,
referenceOutputs: inputs,
);
console.log(result);
`
`
{
'key': 'embedding_similarity',
'score': 0.43,
}
OpenEvals contains some useful prebuilt evaluators for evaluating generated code:
- Type-checking generated code with Pyright and Mypy (Python-only) or TypeScript's built-in type checker (JavaScript only)
- Note that these local type-checking evaluators will not install any dependencies and will ignore errors for these imports
- Sandboxed type-checking and execution evaluators that use E2B to install dependencies and run generated code securely
- LLM-as-a-judge for code
All evaluators in this section accept outputs as a string, an object with a key "messages" that contains a list of messages, or a message-like object with a key "content" that contains a string.
Since LLM outputs with code may contain other text (for example, interleaved explanations with code), OpenEvals code evaluators share some built-in extraction methods for identifying just the code from of LLM outputs.
For any of the evaluators in this section, you can either pass a code_extraction_strategy param set to llm, which will use an llm with a default prompt to directly extract code, or markdown_code_blocks, which will extract anything in markdown code blocks (triple backticks) that is not marked with bash or other shell command languages. If extraction fails for one of these methods, the evaluator response will include a metadata.code_extraction_failed field set to True.
You can alternatively pass a code_extractor param set to a function that takes an LLM output and returns a string of code. The default is to leave the output content untouched ("none").
If using code_extraction_strategy="llm", you can also pass a model string or a client to the evaluator to set which evaluator the model uses for code extraction.code_extractor
If you would like to customize the prompt, you should use the param instead.
For Pyright, you will need to install the pyright CLI on your system:
`bash`
pip install pyright
You can find full installation instructions here.
Then, you can use it as follows:
`python
from openevals.code.pyright import create_pyright_evaluator
evaluator = create_pyright_evaluator()
CODE = """
def sum_of_two_numbers(a, b): return a + b
"""
result = evaluator(outputs=CODE)
print(result)
`
`
{
'key': 'pyright_succeeded',
'score': True,
'comment': None,
}
> [!WARNING]
> The evaluator will ignore reportMissingImports errors. If you want to run type-checking over generated dependencies, check out the sandboxed version of this evaluator.
You can also pass pyright_cli_args to the evaluator to customize the arguments passed to the pyright CLI:
`python`
evaluator = create_pyright_evaluator(
pyright_cli_args=["--flag"]
)
For a full list of supported arguments, see the pyright CLI documentation.
For Mypy, you will need to install mypy on your system:
`bash`
pip install mypy
You can find full installation instructions here.
Then, you can use it as follows:
`python
from openevals.code.mypy import create_mypy_evaluator
evaluator = create_mypy_evaluator()
CODE = """
def sum_of_two_numbers(a, b): return a + b
"""
result = evaluator(outputs=CODE)
print(result)
`
`
{
'key': 'mypy_succeeded',
'score': True,
'comment': None,
}
By default, this evaluator will run with the following arguments:
`
mypy --no-incremental --disallow-untyped-calls --disallow-incomplete-defs --ignore-missing-imports
But you can pass mypy_cli_args to the evaluator to customize the arguments passed to the mypy CLI. This will override the default arguments:
`python`
evaluator = create_mypy_evaluator(
mypy_cli_args=["--flag"]
)
The TypeScript evaluator uses TypeScript's type checker to check the code for correctness.
You will need to install typescript on your system as a dependency (not a dev dependency!):
`bash`
npm install typescript
Then, you can use it as follows (note that you should import from the openevals/code/typescript entrypoint due to the additional required dependency):
`ts
import { createTypeScriptEvaluator } from "openevals/code/typescript";
const evaluator = createTypeScriptEvaluator();
const result = await evaluator({
outputs: "function add(a, b) { return a + b; }",
});
console.log(result);
`
`
{
'key': 'typescript_succeeded',
'score': True,
'comment': None,
}
> [!WARNING]
> The evaluator will ignore reportMissingImports errors. If you want to run type-checking over generated dependencies, check out the sandboxed version of this evaluator.
OpenEvals includes a prebuilt LLM-as-a-judge evaluator for code. The primary differentiator between this one and the more generic LLM-as-judge evaluator is that it will perform the extraction steps detailed above - otherwise it takes the same arguments, including a prompt.
You can run an LLM-as-a-judge evaluator for code as follows:
`ts
import { createCodeLLMAsJudge, CODE_CORRECTNESS_PROMPT } from "openevals";
const evaluator = createCodeLLMAsJudge({
prompt: CODE_CORRECTNESS_PROMPT,
model: "openai:o3-mini",
});
const inputs = Add proper TypeScript types to the following code:
\\\typescript\
function add(a, b) { return a + b; }
\\;
const outputs =
\\\typescript\
function add(a: number, b: number): boolean {
return a + b;
}
\\;
const evalResult = await evaluator({ inputs, outputs });
console.log(evalResult);
`
`
{
"key": "code_correctness",
"score": false,
"comment": "The code has a logical error in its type specification. The function is intended to add two numbers and return their sum, so the return type should be number, not boolean. This mistake makes the solution incorrect according to the rubric. Thus, the score should be: false."
}
LLMs can generate arbitrary code, and if you are running a code evaluator locally, you may not wish to install generated dependencies or run this arbitrary code locally. To solve this, OpenEvals integrates with E2B to run some code evaluators in isolated sandboxes.
Given some output code from an LLM, these sandboxed code evaluators will run scripts in a sandbox that parse out dependencies and install them so that the evaluator has proper context for type-checking or execution.
These evaluators all require a sandbox parameter upon creation, and also accept the code extraction parameters present in the other code evaluators. For Python, there is a special OpenEvalsPython template that includes pyright and uv preinstalled for faster execution, though the evaluator will work with any sandbox.
If you have a custom sandbox with dependencies pre-installed or files already set up, you can supply a sandbox_project_directory (Python) or sandboxProjectDirectory (TypeScript) param when calling the appropriate create method to customize the folder in which type-checking/execution runs.
You can also run Pyright type-checking in an E2B sandbox. The evaluator will run a script to parse out package names
from generated code, then will install those packages in the sandbox and will run Pyright. The evaluator will return any analyzed errors in its comment.
You will need to install the e2b-code-interpreter package, available as an extra:
`bash`
pip install openevals["e2b-code-interpreter"]
Then, you will need to set your E2B API key as an environment variable:
`
export E2B_API_KEY="YOUR_KEY_HERE"
Then, you will need to initialize an E2B sandbox. There is a special OpenEvalsPython template that includes pyright and uv preinstalled for faster execution, though the evaluator will work with any sandbox:
`python
from e2b_code_interpreter import Sandbox
Finally, pass that created sandbox into the
factory function and run it:`python
from openevals.code.e2b.pyright import create_e2b_pyright_evaluatorevaluator = create_e2b_pyright_evaluator(
sandbox=sandbox,
)
CODE = """
from typing import Annotated
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
class State(TypedDict):
messages: Annotated[list, add_messages]
builder = StateGraph(State)
builder.add_node("start", lambda state: state)
builder.compile()
builder.invoke({})
"""
eval_result = evaluator(outputs=CODE)
print(eval_result)
`
{
'key': 'pyright_succeeded',
'score': false,
'comment': '[{"severity": "error", "message": "Cannot access attribute "invoke" for class "StateGraph"...}]',
}
Above, the evaluator identifies and installs the
package inside the sandbox, then runs pyright. The type-check fails because the provided code misuses the imported package, invoking the builder rather than the compiled graph.$3
You can also run TypeScript type-checking in an E2B sandbox. The evaluator will run a script to parse out package names
from generated code, then will install those packages in the sandbox and will run TypeScript. The evaluator will return any analyzed errors in its comment.
You will need to install the official
@e2b/code-interpreter package as a peer dependency:`bash
npm install @e2b/code-interpreter
`Then, you will need to set your E2B API key as an environment variable:
process.env.E2B_API_KEY="YOUR_KEY_HERE"
Next, initialize an E2B sandbox:
ts
import { Sandbox } from "@e2b/code-interpreter";const sandbox = await Sandbox.create();
`And finally, pass the sandbox into the
and run it:`ts
import { createE2BTypeScriptEvaluator } from "openevals/code/e2b";const evaluator = createE2BTypeScriptEvaluator({
sandbox,
});
const CODE =
await StateGraph.invoke({});
const evalResult = await evaluator({ outputs: CODE });
console.log(evalResult);
`
`
{
"key": "typescript_succeeded",
"score": false,
"comment": "(3,18): Property 'invoke' does not exist on type 'typeof StateGraph'."
}
Above, the evaluator identifies and installs @langchain/langgraph, then runs a type-check via TypeScript. The type-check fails because the provided code misuses the imported package.
To further evaluate code correctness, OpenEvals has a sandbox execution evaluator that runs generated code in an E2B sandbox.
The evaluator will run a script to parse out package names from generated code, then will install those packages in the sandbox. The evaluator will then attempt to run the generated code return any analyzed errors in its comment.
You will need to install the official @e2b/code-interpreter package as a peer dependency:
`bash`
npm install @e2b/code-interpreter
Then, you will need to set your E2B API key as an environment variable:
`
process.env.E2B_API_KEY="YOUR_KEY_HERE"
Next, initialize an E2B sandbox:
`ts
import { Sandbox } from "@e2b/code-interpreter";
const sandbox = await Sandbox.create();
`
And finally, pass the sandbox into the create and run it:
`ts
import { createE2BExecutionEvaluator } from "openevals/code/e2b";
const evaluator = createE2BExecutionEvaluator({
sandbox,
});
const CODE =
import { Annotation, StateGraph } from '@langchain/langgraph';
const StateAnnotation = Annotation.Root({
joke: Annotation
topic: Annotation
});
const graph = new StateGraph(StateAnnotation)
.addNode("joke", () => ({}))
.compile();
await graph.invoke({
joke: "foo",
topic: "history",
});;
const evalResult = await evaluator({ outputs });
console.log(evalResult);
`
`
{
"key": "execution_succeeded",
"score": false,
"comment": "file:///home/user/openevals/node_modules/@langchain/langgraph/dist/graph/state.js:197\n throw new Error();\n ^\n\nError: joke is already being used as a state attribute (a.k.a. a channel), cannot also be used as a node name.\n at StateGraph.addNode (/home/user/openevals/node_modules/@langchain/langgraph/src/graph/state.ts:292:13)\n at `
}
Above, the evaluator identifies and installs @langchain/langgraph, then attempts to execute the code. The type-check fails because the provided code misuses the imported package.
If desired, you can pass an environmentVariables object when creating the evaluator. Generated code will have access to these variables within the sandbox, but be cautious, as there is no way to predict exactly what code an LLM will generate.
This package also contains prebuilt evaluators for calculating common metrics such as Levenshtein distance, exact match, etc. You can import and use them as follows:
`ts
import { exactMatch } from "openevals";
const outputs = { a: 1, b: 2 };
const referenceOutputs = { a: 1, b: 2 };
const result = exactMatch(outputs, referenceOutputs);
console.log(result);
`
`
{
key: "equal",
score: true,
}
`ts
import { levenshteinDistance } from "openevals";
const outputs = "The correct answer";
const referenceOutputs = "The correct answer";
const result = levenshteinDistance(outputs, referenceOutputs);
console.log(result);
`
`
{
key: "levenshtein_distance",
score: 0,
}
This evaluator uses LangChain's init_embedding method (for Python) or takes a LangChain embeddings client directly (for TypeScript) and calculates distance between two strings using cosine similarity.
`ts
import { createEmbeddingSimilarityEvaluator } from "openevals";
import { OpenAIEmbeddings } from "@langchain/openai";
const evaluator = createEmbeddingSimilarityEvaluator({
embeddings: new OpenAIEmbeddings({ model: "text-embedding-3-small" }),
});
const result = await evaluator(
outputs: "The weather is nice!",
referenceOutputs: "The weather is very nice!",
);
console.log(result);
`
`
{
key: "embedding_similarity",
score: 0.9147273943905653,
}
If you are building an agent, the evals in this repo are useful for evaluating specific outputs from your agent against references.
However, if you want to get started with more in-depth evals that take into account the entire trajectory of an agent, please check out the agentevals package.
If there are metrics that you want to evaluate that are not covered by any of the above, you can create your own evaluator as well that interacts well with the rest of the openevals ecosystem.
The first thing to note that all evaluators should accept a subset of the following parameters:
- inputs: The inputs to your app.outputs
- : The outputs from your app.reference_outputs
- (Python) or referenceOutputs (TypeScript): The reference outputs to evaluate against.
These parameters can be any value, but should always accept a dict of some kind.
Not all evaluators will use all of these parameters, but they are there to ensure consistency across all evaluators.
Your evaluator may take more parameters as well (e.g. for LLM-as-judge evaluators whose prompts can require additional variables), but for simplicity it's best to stick to the three listed above.
If your evaluator requires additional configuration, you should use a factory function to create your evaluator. These should be named create_ (for example, create_llm_as_judge).
The return values should be a dict (or, if your evaluator evaluates multiple metrics, an list of dicts) with the following keys:
- key: A string representing the name of the metric you are evaluating.score
- : A boolean or number representing the score for the given key.comment
- : A string representing the comment for the given key.
And that's it! Those are the only restrictions.
If you are using LangSmith to track experiments, you should also wrap the internals of your evaluator in the _run_evaluator/_arun_evaluator (Python) or runEvaluator (TypeScript) method. This ensures that the evaluator results are logged to LangSmith properly for supported runners.
This method takes a scorer function as part of its input that returns either:
- A single boolean or number, representing the score for the given key.
- A tuple that contains the score as its first element and a comment justifying the score as its second element.
Here's an example of how you might define a very simple custom evaluator. It only takes into account the outputs of your app and compares them against a regex pattern. It uses a factory function to create the evaluator, since regex is an extra param.
`ts
import { EvaluatorResult } from "openevals/types";
import { _runEvaluator } from "openevals/utils";
/**
* Creates an evaluator that compares the actual output and reference output for similarity by text embedding distance.
* @param {Object} options - The configuration options
* @param {Embeddings} options.embeddings - The embeddings model to use for similarity comparison
* @param {('cosine'|'dot_product')} [options.algorithm='cosine'] - The algorithm to use for embedding similarity
* @returns An evaluator that returns a score representing the embedding similarity
*/
export const createRegexEvaluator = ({
regex,
}: {
regex: RegExp;
}) => {
return async (params: {
outputs: string | Record
}): Promise
const { outputs } = params;
// Tolerate outputs being an object, but convert to string for regex matching
const outputString =
typeof outputs === "string" ? outputs : JSON.stringify(outputs);
const getScore = async (): Promise
return regex.test(outputString);
};
return _runEvaluator(
"regex_match",
getScore,
"regex_match"
);
};
};
`
`ts
const evaluator = createRegexEvaluator({
regex: /some string/,
});
const result = await evaluator({ outputs: "this text contains some string" });
`
`
{
key: "regex_match",
score: true,
}
All openevals evaluators support Python asyncio. As a convention, evaluators that use a factory function will have async put immediately after create_ in the function name (for example, create_async_llm_as_judge), and evaluators used directly will end in async (e.g. exact_match_async).
Here's an example of how to use the create_async_llm_as_judge evaluator asynchronously:
`python
from openevals.llm import create_async_llm_as_judge
evaluator = create_async_llm_as_judge(
prompt="What is the weather in {inputs}?",
model="openai:o3-mini",
)
result = await evaluator(inputs="San Francisco")
`
If you are using the OpenAI client directly, remember to pass in AsyncOpenAI as the judge parameter:
`python
from openai import AsyncOpenAI
evaluator = create_async_llm_as_judge(
prompt="What is the weather in {inputs}?",
judge=AsyncOpenAI(),
model="o3-mini",
)
result = await evaluator(inputs="San Francisco")
`
> [!IMPORTANT]
> The techniques described in this section have changed with the release of 0.1.0. If you are using version 0.0.x of OpenEvals, you can find the old documentation here.
Many LLM applications run across multiple conversation turns with a user. While the LLM-as-judge evaluators in OpenEvals and the trajectory evaluators in AgentEvals are capable of evaluating a full thread of messages, obtaining a representative example thread of messages can be difficult.
To help judge your application's performance over multiple interactions, OpenEvals includes a run_multiturn_simulation method (and its Python async counterpart run_multiturn_simulation_async) for simulating interactions between your app and an end user to help evaluate your app's performance from start to finish.
Here's an example using the OpenAI client directly as a simple chatbot:
`ts
import { OpenAI } from "openai";
import {
createLLMSimulatedUser,
runMultiturnSimulation
createLLMAsJudge,
type ChatCompletionMessage,
} from "openevals";
const client = new OpenAI();
const history = {};
// Your application logic
const app = async ({ inputs, threadId }: { inputs: ChatCompletionMessage, threadId: string }) => {
if (history[threadId] === undefined) {
history[threadId] = [];
}
history[threadId].push(inputs);
const res = await client.chat.completions.create({
model: "gpt-4.1-mini",
messages: [
{
role: "system",
content:
"You are a patient and understanding customer service agent",
},
inputs,
],
});
const responseMessage = res.choices[0].message;
history[threadId].push(responseMessage);
return res.choices[0].message;
};
const user = createLLMSimulatedUser({
system: "You are an aggressive and hostile customer who wants a refund for their car.",
model: "openai:gpt-4.1-mini",
});
const trajectoryEvaluator = createLLMAsJudge({
model: "openai:o3-mini",
prompt: "Based on the below conversation, was the user satisfied?\n{outputs}",
feedbackKey: "satisfaction",
});
const result = await runMultiturnSimulation({
app,
user,
trajectoryEvaluators: [trajectoryEvaluator],
maxTurns: 5,
});
console.log(result);
`
`
{
trajectory: [
{
role: 'user',
content: 'This piece of junk car is a complete disaster! I demand a full refund immediately. How dare you sell me such a worthless vehicle!',
id: 'chatcmpl-BUpXa07LaM7wXbyaNnng1Gtn5Dsbh'
},
{
role: 'assistant',
content: "I'm really sorry to hear about your experience and understand how frustrating this must be. I’d like to help resolve this issue as smoothly as possible. Could you please provide some details about the problem with the vehicle? Once I have more information, I’ll do my best to assist you with a solution, whether it’s a refund or other options. Thank you for your patience.",
refusal: null,
annotations: [],
id: 'd7520f6a-7cf8-46f8-abe4-7df04f134482'
},
...
{
role: 'assistant',
content: "I truly understand your frustration and sincerely apologize for the inconvenience you've experienced. I want to resolve this issue for you as quickly as possible. \n" +
'\n' +
'Please allow me a moment to review your case, and I will do everything I can to expedite your refund. Your patience is greatly appreciated, and I am committed to resolving this matter to your satisfaction.',
refusal: null,
annotations: [],
id: 'a0536d4f-9353-4cfa-84df-51c8d29e076d'
}
],
evaluatorResults: [
{
key: 'satisfaction',
score: false,
comment: 'The user is clearly dissatisfied and expresses frustration throughout the conversation. Their repeated demands for a refund and threats to escalate the situation indicate a lack of satisfaction with the responses provided. They specifically mention they don’t want excuses or further delays, highlighting their dissatisfaction with the service. Thus, the score should be: false.',
metadata: undefined
}
]
}
There are two main components:
- app: Your application, or a function wrapping it. Must accept a chat message (dict with "role" and "content" keys) as an input arg and a thread_id as a kwarg. Should accept other kwargs as more may be added in future releases. Returns a chat message as output with at least role and content keys.app
- Note that your will only receive the next message from the simulated user as input, and therefore should statefully track the current history internally based on thread_id if needed.user
- : The simulated user. Must accept the current trajectory as a list of messages as an input arg and kwargs for thread_id and turn_counter. Should accept other kwargs as more may be added in future releases. Returns a chat message as output. May also be a list of string or message responses.create_llm_simulated_user
- In the example above, this is an imported prebuilt function named which uses an LLM to generate user responses, though you are free to define your own function as well. See this section for more information.
The simulation will call the user first to obtain the first input for app, which should return a chat message. The returned message is passed back into user, and so on until the simulator reaches max_turns or an optionally passed stopping_condition returns True.
The returned messages are deduped by id and added to an internal list of messages representing a trajectory, which is returned as part of the simulator results. If a returned message does not contain an id field, the simulator will automatically generate one.
The other accepted parameters are as follows:
- thread_id/threadId: An optional thread id that identifies the current interaction, used by your app to load state. Will default to a UUID if not provided.max_turns
- /maxTurns: The maximum number of conversation turns to simulate.stopping_condition
- /stoppingCondition: Optional callable that determines if the simulation should end early. Takes the current trajectory as a list of messages as an input arg and a kwarg named turn_counter, and should return a boolean.trajectory_evaluators
- /trajectoryEvaluators: Optional evaluators that run at the end of the simulation. These will receive the final trajectory as a kwarg named outputs.reference_outputs
- /referenceOutputs: An optional reference trajectory which will be passed directly through to the provided trajectory_evaluators.
You must pass at least one of max_turns or stopping_condition. Once one of these triggers, the final trajectory will be passed to provided trajectory evaluators, which will receive the final trajectory as an "outputs" kwarg.
The simulator itself is not an evaluator and will not return or log any feedback. Instead, it will return a MultiturnSimulationResult with the following structure:
`ts`
type MultiturnSimulationResult = {
evaluatorResults: EvaluatorResult[];
trajectory: ChatCompletionMessage[];
};
Where evaluator_results/evaluatorResults are the results from the passed trajectory_evaluators and trajectory is the final trajectory.
The Python async version works the same way, but requires async functions to be passed rather than sync ones.
The user parameter is a function that accepts the current trajectory (and a thread_id/threadId kwarg), then returns a message with role="user" that will be passed back to your app. We suggest starting with the prebuilt method returned by create_llm_simulated_user, but you can also customize your own if desired.
> [!NOTE]
> The simulated user is pretending to be a human, and should therefore return a user message, not an assistant message!
OpenEvals includes a prebuilt create_llm_simulated_user method that uses an LLM to take on the role of a user and generate responses based on a system prompt:
`ts
import { createLLMSimulatedUser } from "openevals";
const user = createLLMSimulatedUser({
system: "You are an aggressive and hostile customer who wants a refund for their car.",
model: "openai:gpt-4.1-mini",
});
`
You can also pass an array of fixed_responses`, which the simulated user will return in order. Here is an example of a simulated user set u