@dooor-ai/toolkit/dist/evals/faithfulness.js

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

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.FaithfulnessEval = void 0;
const base_1 = require("./base");
const cortexdb_client_1 = require("../observability/cortexdb-client");
/**
 * FaithfulnessEval - Measures if the LLM's answer is grounded in the provided context
 *
 * Critical for RAG systems to prevent hallucination. Checks if claims in the answer
 * are supported by the retrieved context.
 *
 * Example:
 * ```typescript
 * const eval = new FaithfulnessEval({
 *   threshold: 0.8,
 *   context: "The Eiffel Tower is 330 meters tall."
 * });
 * const result = await eval.evaluate(
 *   "How tall is the Eiffel Tower?",
 *   "The Eiffel Tower is 330 meters tall."
 * );
 * // result.score = 1.0, result.passed = true
 * ```
 */
class FaithfulnessEval extends base_1.Eval {
    constructor(config = {}) {
        super(config);
        this.context = config.context;
    }
    get name() {
        return "FaithfulnessEval";
    }
    /**
     * Set context dynamically (useful for RAG pipelines)
     */
    setContext(context) {
        this.context = context;
    }
    async evaluate(input, output, metadata) {
        const startTime = Date.now();
        // Get context from metadata if not set directly
        const context = this.context || metadata?.context || metadata?.retrievedDocs;
        if (!context) {
            return {
                name: this.name,
                score: 0.5,
                passed: false,
                details: "No context provided for faithfulness evaluation. Pass context via config or metadata.",
                metadata: {
                    latency: Date.now() - startTime,
                },
                timestamp: new Date(),
            };
        }
        try {
            const cortexClient = (0, cortexdb_client_1.getCortexDBClient)();
            const providerName = (0, cortexdb_client_1.getGlobalProviderName)();
            const prompt = this.buildPrompt(input, output, context);
            const response = await cortexClient.invokeAI({
                prompt,
                usage: "evaluation",
                providerName: providerName || undefined,
                temperature: 0.0,
                maxTokens: 300,
            });
            const score = this.parseScore(response.text);
            const passed = score >= this.getThreshold();
            return {
                name: this.name,
                score,
                passed,
                details: `Faithfulness score: ${score.toFixed(2)}. ${passed ? "PASSED" : "FAILED"} (threshold: ${this.getThreshold()})`,
                metadata: {
                    latency: Date.now() - startTime,
                    judgeResponse: response.text,
                    contextLength: context.length,
                },
                timestamp: new Date(),
            };
        }
        catch (error) {
            console.error("FaithfulnessEval 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(),
            };
        }
    }
    buildPrompt(question, answer, context) {
        return `You are an expert evaluator. Your task is to assess if the answer is faithful to (grounded in) the provided context.

Context: """
${context}
"""

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

Evaluate faithfulness:
- 1.0 = All claims in answer are directly supported by context
- 0.7-0.9 = Most claims supported, minor unsupported details
- 0.4-0.6 = Some claims supported, some not found in context
- 0.0-0.3 = Answer contains hallucinations or unsupported claims

Output ONLY a JSON object in this exact format:
{
  "score": 0.9,
  "reasoning": "Brief explanation of what's supported and what's not"
}`;
    }
    parseScore(response) {
        try {
            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));
                }
            }
            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;
        }
        catch (error) {
            console.error("Error parsing score:", error);
            return 0.5;
        }
    }
}
exports.FaithfulnessEval = FaithfulnessEval;