@dooor-ai/toolkit/src/evals/answer-relevancy.ts

Guards, Evals & Observability for AI applications - works seamlessly with LangChain/LangGraph

import { Eval } from "./base";
import { EvalResult, EvalConfig } from "../core/types";
import { getCortexDBClient, getGlobalProviderName } from "../observability/cortexdb-client";

/**
 * AnswerRelevancyEval - Measures if the LLM's answer is relevant to the question
 *
 * Uses LLM-as-judge to score relevancy on a 0-1 scale.
 *
 * Example:
 * ```typescript
 * const eval = new AnswerRelevancyEval({ threshold: 0.7 });
 * const result = await eval.evaluate(
 *   "What is the capital of France?",
 *   "The capital of France is Paris."
 * );
 * // result.score = 0.95, result.passed = true
 * ```
 */
export class AnswerRelevancyEval extends Eval {
  get name(): string {
    return "AnswerRelevancyEval";
  }

  async evaluate(
    input: string,
    output: string,
    metadata?: Record<string, any>
  ): Promise<EvalResult> {
    const startTime = Date.now();

    try {
      // Use CortexDB AI proxy for evaluation
      const cortexClient = getCortexDBClient();
      const providerName = getGlobalProviderName();

      const prompt = this.buildPrompt(input, output);

      const response = await cortexClient.invokeAI({
        prompt,
        usage: "evaluation",
        providerName: providerName || undefined,
        temperature: 0.0, // Deterministic for consistency
        maxTokens: 200,
      });

      const score = this.parseScore(response.text);
      const passed = score >= this.getThreshold();

      return {
        name: this.name,
        score,
        passed,
        details: `Answer relevancy score: ${score.toFixed(2)}. ${passed ? "PASSED" : "FAILED"} (threshold: ${this.getThreshold()})`,
        metadata: {
          latency: Date.now() - startTime,
          judgeResponse: response.text,
        },
        timestamp: new Date(),
      };
    } catch (error) {
      console.error("AnswerRelevancyEval failed:", error);

      return {
        name: this.name,
        score: 0.5,
        passed: false,
        details: `Eval failed: ${error instanceof Error ? error.message : "Unknown error"}`,
        metadata: {
          error: String(error),
          latency: Date.now() - startTime,
        },
        timestamp: new Date(),
      };
    }
  }

  private buildPrompt(question: string, answer: string): string {
    return `You are an expert evaluator. Your task is to assess if the answer is relevant to the question.

Question: "${question}"
Answer: "${answer}"

Rate the relevancy on a scale of 0.0 to 1.0:
- 1.0 = Perfectly relevant, directly addresses the question
- 0.7-0.9 = Mostly relevant, addresses core question with some tangents
- 0.4-0.6 = Partially relevant, touches on topic but misses key points
- 0.0-0.3 = Not relevant, does not address the question

Output ONLY a JSON object in this exact format:
{
  "score": 0.85,
  "reasoning": "Brief explanation of the score"
}`;
  }

  private parseScore(response: string): number {
    try {
      // Try to extract JSON from response
      const jsonMatch = response.match(/\{[\s\S]*\}/);
      if (jsonMatch) {
        const parsed = JSON.parse(jsonMatch[0]);
        if (typeof parsed.score === "number") {
          return Math.max(0, Math.min(1, parsed.score));
        }
      }

      // Fallback: look for decimal number
      const numberMatch = response.match(/\b0?\.\d+\b|\b1\.0\b|\b[01]\b/);
      if (numberMatch) {
        return Math.max(0, Math.min(1, parseFloat(numberMatch[0])));
      }

      console.warn("Could not parse score from response:", response);
      return 0.5; // Default middle score on parse failure
    } catch (error) {
      console.error("Error parsing score:", error);
      return 0.5;
    }
  }
}