@dooor-ai/toolkit/src/evals/summarization.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 SummarizationConfig extends EvalConfig {
  /** Original text that was summarized */
  originalText?: string;
}

/**
 * SummarizationEval - Evaluates quality of text summarization
 *
 * Checks if summary captures key points, maintains accuracy, and is concise.
 *
 * Example:
 * ```typescript
 * const eval = new SummarizationEval({
 *   threshold: 0.8,
 *   originalText: "The Eiffel Tower is a wrought-iron lattice tower on the Champ de Mars in Paris, France. It is named after the engineer Gustave Eiffel, whose company designed and built the tower from 1887 to 1889."
 * });
 * const result = await eval.evaluate(
 *   "Summarize this text.",
 *   "The Eiffel Tower in Paris was designed by Gustave Eiffel and built 1887-1889."
 * );
 * // result.score = 0.9 (good summary), result.passed = true
 * ```
 */
export class SummarizationEval extends Eval {
  private originalText?: string;

  constructor(config: SummarizationConfig = {}) {
    super(config);
    this.originalText = config.originalText;
  }

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

  /**
   * Set original text dynamically
   */
  setOriginalText(text: string): void {
    this.originalText = text;
  }

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

    const original = this.originalText || metadata?.originalText || metadata?.sourceText;

    if (!original) {
      return {
        name: this.name,
        score: 0.5,
        passed: false,
        details: "No original text provided for summarization evaluation. Pass 'originalText' via config or metadata.",
        metadata: {
          latency: Date.now() - startTime,
        },
        timestamp: new Date(),
      };
    }

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

      const prompt = this.buildPrompt(original, 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: `Summarization quality score: ${score.toFixed(2)}. ${passed ? "PASSED" : "FAILED"} (threshold: ${this.getThreshold()})`,
        metadata: {
          latency: Date.now() - startTime,
          judgeResponse: response.text,
          originalLength: original.length,
          summaryLength: output.length,
          compressionRatio: (output.length / original.length).toFixed(2),
        },
        timestamp: new Date(),
      };
    } catch (error) {
      console.error("SummarizationEval 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(originalText: string, summary: string): string {
    return `You are an expert evaluator. Your task is to evaluate the quality of a SUMMARY.

Original Text: """
${originalText}
"""

Summary: """
${summary}
"""

Evaluate summarization quality based on:
1. **Coverage**: Does it capture the main points?
2. **Accuracy**: Is it factually correct?
3. **Conciseness**: Is it appropriately brief?
4. **Coherence**: Is it well-structured and readable?

Score:
- 1.0 = Excellent summary, captures all key points accurately and concisely
- 0.7-0.9 = Good summary, minor omissions or verbosity
- 0.4-0.6 = Mediocre, misses important points or too verbose
- 0.0-0.3 = Poor, inaccurate or misses most key information

Output ONLY a JSON object in this exact format:
{
  "score": 0.85,
  "reasoning": "Assessment of coverage, accuracy, conciseness, and coherence"
}`;
  }

  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;
    }
  }
}