// KnowledgeDetector（veridrop detectors/knowledge.py）：critical 题一题一个请求；coverage 题并成一个请求，按行号拆答案。
// 命中 expected 关键词、且不含 anti 关键词算对。

import { KNOWLEDGE_QUESTIONS } from "../data/knowledge_questions.js";
import type { DetectorResult } from "../models.js";
import { cpSlice, joinText, splitlines } from "../py.js";
import { ActiveDetector, errText, type DetectClient } from "./base.js";

const MAX_TOKENS_CRITICAL = 60;
const MAX_TOKENS_COVERAGE = 400;

export interface Question {
  id: string;
  type: string;
  prompt: string;
  expected_keywords?: readonly string[];
  expected_keyword_match?: string;
  anti_keywords?: readonly string[];
  applicable_models?: readonly string[];
  expected_by_model?: Readonly<Record<string, readonly string[]>>;
}

function applies(q: Question, model: string): boolean {
  const allow = q.applicable_models;
  if (!allow || allow.length === 0) return true;
  return allow.some((m) => model.startsWith(m) || m.startsWith(model));
}

export function buildCombinedPrompt(coverage: readonly Question[]): string {
  const lines = [
    `Please answer these ${coverage.length} questions briefly. ` +
      "Reply with one short answer per line, prefixed by question number " +
      "(e.g. '1. <answer>'). If you don't know an answer, reply 'unknown' " +
      "for that line — do not guess.",
  ];
  coverage.forEach((q, i) => lines.push(`${i + 1}. ${q.prompt}`));
  return lines.join("\n\n");
}

// "1. answer"、"2) answer"、"3: answer"、"  4 - answer" 都认。
const LINE_RE = /^\s*(\d+)\s*[.):\-]\s*(.+?)\s*$/u;

export function parseNumberedAnswers(text: string, n: number): Map<number, string> {
  const out = new Map<number, string>();
  for (const line of splitlines(text)) {
    const m = LINE_RE.exec(line);
    if (!m) continue;
    const idx = Number.parseInt(m[1]!, 10);
    if (idx >= 1 && idx <= n) out.set(idx, m[2]!);
  }
  return out;
}

export function grade(answer: string, q: Question, model: string): boolean {
  if (!answer) return false;
  const a = answer.toLowerCase();
  const anti = (q.anti_keywords ?? []).map((k) => k.toLowerCase());
  if (anti.some((k) => a.includes(k))) return false;
  let expected: string[] | null = null;
  for (const [key, kws] of Object.entries(q.expected_by_model ?? {})) {
    if (model.startsWith(key) || key.startsWith(model)) {
      expected = kws.map((k) => k.toLowerCase());
      break;
    }
  }
  if (expected === null) expected = (q.expected_keywords ?? []).map((k) => k.toLowerCase());
  if (expected.length === 0) return true;
  if ((q.expected_keyword_match ?? "all") === "any") return expected.some((k) => a.includes(k));
  return expected.every((k) => a.includes(k));
}

export class KnowledgeDetector extends ActiveDetector {
  readonly name = "knowledge";
  readonly displayName = "知识准确度";
  readonly weight = 10;

  async run(client: DetectClient, model: string): Promise<DetectorResult> {
    const questions = KNOWLEDGE_QUESTIONS.questions as readonly Question[];
    const applicable = questions.filter((q) => applies(q, model));
    if (applicable.length === 0) return this.result("skip", 0, { reason: "no applicable questions for model" });

    const critical = applicable.filter((q) => q.type === "critical");
    const coverage = applicable.filter((q) => q.type === "coverage");
    const perQuestion: Record<string, unknown>[] = [];

    for (const q of critical) {
      let resp: Record<string, unknown>;
      try {
        ({ resp } = await client.messagesCreate({
          model,
          max_tokens: MAX_TOKENS_CRITICAL,
          temperature: 0,
          messages: [{ role: "user", content: q.prompt }],
        }));
      } catch (e) {
        perQuestion.push({ id: q.id, passed: false, error: errText(e), answer: "" });
        continue;
      }
      const answer = joinText(resp["content"]);
      perQuestion.push({ id: q.id, passed: grade(answer, q, model), answer: cpSlice(answer, 200) });
    }

    if (coverage.length > 0) {
      let answerText: string | null = null;
      try {
        const { resp } = await client.messagesCreate({
          model,
          max_tokens: MAX_TOKENS_COVERAGE,
          temperature: 0,
          messages: [{ role: "user", content: buildCombinedPrompt(coverage) }],
        });
        answerText = joinText(resp["content"]);
      } catch (e) {
        for (const q of coverage) perQuestion.push({ id: q.id, passed: false, error: errText(e), answer: "" });
      }
      if (answerText !== null) {
        const parsed = parseNumberedAnswers(answerText, coverage.length);
        coverage.forEach((q, i) => {
          const a = parsed.get(i + 1) ?? "";
          perQuestion.push({ id: q.id, passed: grade(a, q, model), answer: cpSlice(a, 200) });
        });
      }
    }

    const passes = perQuestion.filter((r) => r["passed"]).length;
    const total = perQuestion.length;
    const score = total ? (passes / total) * 100 : 0;
    return this.result(score >= 70 ? "pass" : "fail", score, {
      passes,
      total,
      per_question: perQuestion,
      request_count: critical.length + (coverage.length > 0 ? 1 : 0),
    });
  }
}