Cosine Similarity

Measure similarity between generated and reference texts. Improve alignment with semantic scoring.

The Cosine Similarity metric is a span-level metric used to measure the similarity between the response and the ground truth in an Agentic application. This metric is ideal for evaluating how closely a model's response aligns with the expected output.

---

Schema

The schema for the Cosine Similarity metric includes:

• Ground Truth: The expected output for the given input.

• Response: The model's actual output.

---

How to Run the Metric

1. Access the Dataset

   • Navigate to the dataset where you want to evaluate responses.

   • Click on the Evaluate button.

2. Select the Metric

   • Choose Cosine Similarity-Alteryx from the list of metrics.

   • You can rename the metric for clarity, if needed.

3. Choose the Evaluation Type

   • Select the component type you wish to evaluate:

     • LLM: Evaluate spans related to language model outputs.

     • Agent: Evaluate agent-level interactions.

     • Tool: Evaluate tool-related outputs.

4. Define the Schema

   • Specify the span name you want to evaluate.

   • Choose the parameter for evaluation, such as:

     • input

     • output

     • ground truth

5. Configure the Model

   • Select the model configuration to be used for evaluation.

6. Set Passing Criteria

   • Define the pass/fail threshold for the metric.

7. Run the Metric

   • Click on Run to start the evaluation process.

---

When to Run the Cosine Similarity Metric?

• To Evaluate Response Accuracy: Use this metric when you want to determine how closely the system's response matches the expected ground truth.

• Span-Level Analysis: Cosine Similarity is particularly useful for evaluating specific spans or segments of a trace where text similarity is critical.

• For Model Comparison: Employ this metric to compare outputs from different models or configurations and identify which aligns better with the ground truth.