@dooor-ai/toolkit/src/evals/knowledge-retention.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 KnowledgeRetentionConfig extends EvalConfig {
  /** Conversation history (previous messages) */
  conversationHistory?: Array<{ role: string; content: string }>;
}

/**
 * KnowledgeRetentionEval - Measures if LLM remembers previous conversation context
 *
 * Evaluates whether the response demonstrates awareness of earlier messages,
 * important for multi-turn conversations and chatbots.
 *
 * Example:
 * ```typescript
 * const eval = new KnowledgeRetentionEval({
 *   threshold: 0.8,
 *   conversationHistory: [
 *     { role: "user", content: "My name is Alice." },
 *     { role: "assistant", content: "Nice to meet you, Alice!" }
 *   ]
 * });
 * const result = await eval.evaluate(
 *   "What's my name?",
 *   "Your name is Alice."
 * );
 * // result.score = 1.0, result.passed = true
 * ```
 */
export class KnowledgeRetentionEval extends Eval {
  private conversationHistory?: Array<{ role: string; content: string }>;

  constructor(config: KnowledgeRetentionConfig = {}) {
    super(config);
    this.conversationHistory = config.conversationHistory;
  }

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

  /**
   * Set conversation history dynamically
   */
  setConversationHistory(history: Array<{ role: string; content: string }>): void {
    this.conversationHistory = history;
  }

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

    const history = this.conversationHistory || metadata?.conversationHistory || metadata?.history;

    if (!history || history.length === 0) {
      return {
        name: this.name,
        score: 0.5,
        passed: false,
        details: "No conversation history provided. Pass 'conversationHistory' via config or metadata.",
        metadata: {
          latency: Date.now() - startTime,
        },
        timestamp: new Date(),
      };
    }

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

      const prompt = this.buildPrompt(history, 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: `Knowledge retention score: ${score.toFixed(2)}. ${passed ? "PASSED" : "FAILED"} (threshold: ${this.getThreshold()})`,
        metadata: {
          latency: Date.now() - startTime,
          judgeResponse: response.text,
          historyLength: history.length,
        },
        timestamp: new Date(),
      };
    } catch (error) {
      console.error("KnowledgeRetentionEval 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(
    history: Array<{ role: string; content: string }>,
    currentInput: string,
    currentOutput: string
  ): string {
    const historyText = history
      .map((msg) => `${msg.role.toUpperCase()}: ${msg.content}`)
      .join("\n");

    return `You are an expert evaluator. Your task is to assess if the response demonstrates KNOWLEDGE RETENTION from previous conversation.

Conversation History:
${historyText}

Current Turn:
USER: ${currentInput}
ASSISTANT: ${currentOutput}

Evaluate knowledge retention:
- 1.0 = Perfect retention, references earlier context appropriately
- 0.7-0.9 = Good retention, mostly aware of history
- 0.4-0.6 = Partial retention, misses some important context
- 0.0-0.3 = Poor retention, ignores or contradicts earlier information

Output ONLY a JSON object in this exact format:
{
  "score": 0.9,
  "reasoning": "Assessment of how well the response uses conversation history"
}`;
  }

  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;
    }
  }
}