Completeness

Understand Galileo's Completeness Metric

The metric is mainly intended for RAG workflows.

Definition: Measures how thoroughly your model's response covered the relevant information available in the context provided.

Completeness and Context Adherence are closely related, and designed to complement one another:

• Context Adherence answers the question, "is the model's response consistent with the information in the context?"

• Completeness answers the question, "is the relevant information in the context fully reflected in the model's response?"

In other words, if Context Adherence is "precision," then Completeness is "recall."

Consider this simple, stylized example that illustrates the distinction:

• User query: "Who was Galileo Galilei?"

• Context: "Galileo Galilei was an Italian astronomer."

• Model response: "Galileo Galilei was Italian."

This response would receive a perfect Context Adherence score: everything the model said is supported by the context.

But this is not an ideal response. The context also specified that Galileo was an astronomer, and the user probably wants to know that information as well.

Hence, this response would receive a low Completeness score. Tracking Completeness alongside Context Adherence allows you to detect cases like this one, where the model is "too reticent" and fails to mention relevant information.

Usefulness: To fix low Completeness values, we recommend adjusting the prompt to tell the model to include all the relevant information it can find in the provided context.

Basic vs Plus

We offer two ways of calculating Completeness: Basic and Plus.

Completeness Basic is computed using Galileo in-house small language models. They're free of cost, but lack 'explanations'. Completeness Basic is a cost effective way to scale up you RAG evaluation workflows.

Completeness Plus is computed using the Chainpoll technique. It relies on OpenAI models so it incurs an additional cost. Completeness Plus has shown better results in internal benchmarks. Additionally, Plus offers explanations for its ratings (i.e. why a response was or was not complete).