@dooor-ai/toolkit/dist/evals/contextual-precision.js

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

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.ContextualPrecisionEval = void 0;
const base_1 = require("./base");
const cortexdb_client_1 = require("../observability/cortexdb-client");
/**
 * ContextualPrecisionEval - Measures if the retrieved context is relevant (no noise)
 *
 * Evaluates whether the retrieved documents are actually useful for answering
 * the question. High precision means low noise, no irrelevant docs.
 *
 * Example:
 * ```typescript
 * const eval = new ContextualPrecisionEval({
 *   threshold: 0.8,
 *   context: "Paris is the capital of France. The weather is sunny today."
 * });
 * const result = await eval.evaluate(
 *   "What is the capital of France?",
 *   "Paris"
 * );
 * // result.score = 0.75 (50% relevant, 50% noise), result.passed = false
 * ```
 */
class ContextualPrecisionEval extends base_1.Eval {
    constructor(config = {}) {
        super(config);
        this.context = config.context;
    }
    get name() {
        return "ContextualPrecisionEval";
    }
    /**
     * Set context dynamically
     */
    setContext(context) {
        this.context = context;
    }
    async evaluate(input, output, metadata) {
        const startTime = Date.now();
        const context = this.context || metadata?.context || metadata?.retrievedDocs;
        if (!context) {
            return {
                name: this.name,
                score: 0.5,
                passed: false,
                details: "No context provided for precision 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, 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: `Contextual precision 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("ContextualPrecisionEval 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, context) {
        return `You are an expert evaluator. Your task is to measure PRECISION: how much of the retrieved context is actually relevant to the question?

Question: "${question}"

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

Evaluate precision (relevance ratio):
- 1.0 = 100% of context is relevant, zero noise
- 0.7-0.9 = Most context is relevant, minor irrelevant parts
- 0.4-0.6 = Mixed, significant irrelevant content
- 0.0-0.3 = Mostly irrelevant, high noise

Output ONLY a JSON object in this exact format:
{
  "score": 0.85,
  "reasoning": "What percentage is relevant vs irrelevant"
}`;
    }
    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.ContextualPrecisionEval = ContextualPrecisionEval;