@dooor-ai/toolkit/src/evals/hallucination.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";

export interface HallucinationConfig extends EvalConfig {
  /** Context to check against (optional, otherwise checks for factual coherence) */
  context?: string;
}

/**
 * HallucinationEval - Detects when LLM makes up facts or unsupported claims
 *
 * If context is provided, checks if answer contains claims not in context.
 * If no context, evaluates internal consistency and factual plausibility.
 *
 * Example:
 * ```typescript
 * const eval = new HallucinationEval({
 *   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 450 meters tall and was built in 1950."
 * );
 * // result.score = 0.2 (high hallucination), result.passed = false
 * ```
 */
export class HallucinationEval extends Eval {
  private context?: string;

  constructor(config: HallucinationConfig = {}) {
    super(config);
    this.context = config.context;
  }

  get name(): string {
    return "HallucinationEval";
  }

  /**
   * Set context dynamically
   */
  setContext(context: string): void {
    this.context = context;
  }

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

    const context = this.context || metadata?.context || metadata?.retrievedDocs;

    try {
      const cortexClient = getCortexDBClient();
      const providerName = getGlobalProviderName();

      const prompt = context
        ? this.buildPromptWithContext(input, output, context)
        : this.buildPromptWithoutContext(input, output);

      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: `Hallucination score: ${score.toFixed(2)} (higher is better). ${passed ? "PASSED" : "FAILED"} (threshold: ${this.getThreshold()})`,
        metadata: {
          latency: Date.now() - startTime,
          judgeResponse: response.text,
          hasContext: !!context,
        },
        timestamp: new Date(),
      };
    } catch (error) {
      console.error("HallucinationEval 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 buildPromptWithContext(question: string, answer: string, context: string): string {
    return `You are an expert evaluator. Your task is to detect HALLUCINATIONS: claims in the answer that are NOT supported by the context.

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

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

Evaluate hallucination (inverse scale):
- 1.0 = Zero hallucinations, all claims are grounded
- 0.7-0.9 = Minor unsupported details
- 0.4-0.6 = Some significant hallucinations
- 0.0-0.3 = Major hallucinations, answer is mostly fabricated

Output ONLY a JSON object in this exact format:
{
  "score": 0.9,
  "reasoning": "What claims are supported vs hallucinated"
}`;
  }

  private buildPromptWithoutContext(question: string, answer: string): string {
    return `You are an expert evaluator. Your task is to detect HALLUCINATIONS: claims in the answer that are factually implausible or internally inconsistent.

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

Evaluate factual plausibility:
- 1.0 = All claims seem factually plausible and consistent
- 0.7-0.9 = Mostly plausible, minor inconsistencies
- 0.4-0.6 = Some implausible or contradictory claims
- 0.0-0.3 = Major factual errors or contradictions

Output ONLY a JSON object in this exact format:
{
  "score": 0.85,
  "reasoning": "Assessment of factual plausibility"
}`;
  }

  private parseScore(response: string): number {
    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;
    }
  }
}