|
| 1 | +package org.springframework.ai.evaluation; |
| 2 | + |
| 3 | +import java.util.Collections; |
| 4 | + |
| 5 | +import org.springframework.ai.chat.client.ChatClient; |
| 6 | + |
| 7 | +/** |
| 8 | + * Evaluates the correctness of a generated answer. |
| 9 | + * |
| 10 | + * The evaluator relies on a reference answer to judge the correctness of the generated |
| 11 | + * answer. |
| 12 | + * |
| 13 | + * The evaluation response includes a score between 1 and 5, where 1 is the worst and 5 is |
| 14 | + * the best. The evaluator also provides reasoning for the score. |
| 15 | + * |
| 16 | + * Passing is determined by the score being greater than or equal to a threshold. |
| 17 | + * |
| 18 | + * @author Craig Walls |
| 19 | + * @since 1.0.0 M1 |
| 20 | + */ |
| 21 | +public class CorrectnessEvaluator implements Evaluator { |
| 22 | + |
| 23 | + private static final String DEFAULT_REFERENCE_ANSWER = "(NO REFERENCE ANSWER SUPPLIED)"; |
| 24 | + |
| 25 | + private static final String DEFAULT_SYSTEM_PROMPT_TEXT = """ |
| 26 | + You are an expert evaluation system for a question answering chatbot. |
| 27 | +
|
| 28 | + You are given the following information: |
| 29 | + - a user query, and |
| 30 | + - a generated answer |
| 31 | +
|
| 32 | + You may also be given a reference answer to use for reference in your evaluation. |
| 33 | +
|
| 34 | + Your job is to judge the relevance and correctness of the generated answer. |
| 35 | + Output a single score that represents a holistic evaluation. |
| 36 | + You must return your response in a line with only the score. |
| 37 | + Do not return answers in any other format. |
| 38 | + On a separate line provide your reasoning for the score as well. |
| 39 | +
|
| 40 | + Follow these guidelines for scoring: |
| 41 | + - Your score has to be between 1 and 5, where 1 is the worst and 5 is the best. |
| 42 | + - If the generated answer is not relevant to the user query, |
| 43 | + you should give a score of 1. |
| 44 | + - If the generated answer is relevant but contains mistakes, |
| 45 | + you should give a score between 2 and 3. |
| 46 | + - If the generated answer is relevant and fully correct, |
| 47 | + you should give a score between 4 and 5. |
| 48 | +
|
| 49 | + Example Response: |
| 50 | + 4.0 |
| 51 | + The generated answer has the exact same metrics as the reference answer, |
| 52 | + but it is not as concise. |
| 53 | + """; |
| 54 | + |
| 55 | + private static final String DEFAULT_USER_PROMPT_TEMPLATE = """ |
| 56 | + ## User Query |
| 57 | + {query} |
| 58 | +
|
| 59 | + ## Reference Answer |
| 60 | + {reference_answer} |
| 61 | +
|
| 62 | + ## Generated Answer |
| 63 | + {generated_answer} |
| 64 | + """; |
| 65 | + |
| 66 | + private final ChatClient.Builder chatClientBuilder; |
| 67 | + |
| 68 | + private float scoreThreshold = 4.0f; |
| 69 | + |
| 70 | + public CorrectnessEvaluator(ChatClient.Builder chatClientBuilder, float scoreThreshold) { |
| 71 | + this.chatClientBuilder = chatClientBuilder; |
| 72 | + this.scoreThreshold = scoreThreshold; |
| 73 | + } |
| 74 | + |
| 75 | + @Override |
| 76 | + public EvaluationResponse evaluate(EvaluationRequest evaluationRequest) { |
| 77 | + final String referenceAnswer = (evaluationRequest instanceof CorrectnessEvaluationRequest) |
| 78 | + ? ((CorrectnessEvaluationRequest) evaluationRequest).getReferenceAnswer() : DEFAULT_REFERENCE_ANSWER; |
| 79 | + |
| 80 | + var query = evaluationRequest.getUserText(); |
| 81 | + var generatedAnswer = evaluationRequest.getResponseContent(); |
| 82 | + |
| 83 | + CorrectnessEvaluation evaluationResult = this.chatClientBuilder.build() |
| 84 | + .prompt() |
| 85 | + .system(systemSpec -> systemSpec.text(DEFAULT_SYSTEM_PROMPT_TEXT)) |
| 86 | + .user(userSpec -> userSpec.text(DEFAULT_USER_PROMPT_TEMPLATE) |
| 87 | + .param("query", query) |
| 88 | + .param("reference_answer", referenceAnswer) |
| 89 | + .param("generated_answer", generatedAnswer)) |
| 90 | + .call() |
| 91 | + .entity(CorrectnessEvaluation.class); |
| 92 | + |
| 93 | + boolean passing = evaluationResult.score() >= this.scoreThreshold; |
| 94 | + |
| 95 | + return new EvaluationResponse(passing, evaluationResult.score(), evaluationResult.reasoning(), |
| 96 | + Collections.emptyMap()); |
| 97 | + } |
| 98 | + |
| 99 | + private record CorrectnessEvaluation(float score, String reasoning) { |
| 100 | + } |
| 101 | + |
| 102 | +} |
0 commit comments