export interface CriticCardInput {
  id: string;
  url: string;
  title: string;
  content: string;
  engine?: string;
  relevanceScore?: number;
  sourceTier?: string;
}

export interface FastCriticPartition {
  autoApproved: Array<{
    id: string;
    tier: "A" | "B";
    confidence: "High" | "Medium";
    reason: string;
  }>;
  autoRejected: Array<{
    id: string;
    reason: string;
  }>;
  needsLlmGrading: Array<CriticCardInput & { compactExcerpt: string }>;
}

const JUNK_TITLE_OR_URL_REGEX =
  /\b(merriam-webster\.com\/dictionary|dictionary\.cambridge\.org|collinsdictionary\.com|vocabulary\.com|\/search\?|search\s+page|login|sign\s+in)\b/i;

const CANONICAL_ENGINES = new Set([
  "openalex",
  "crossref",
  "arxiv",
  "pubmed",
  "wikipedia",
]);

/**
 * Partitions evidence cards BEFORE calling the LLM Critic:
 * 1. Rejects dictionary/search-page noise ("Search Page", "merriam-webster.com/dictionary/original").
 * 2. Auto-approves Tier-0 canonical academic/reference cards & high-relevance (rel >= 0.85) cards.
 * 3. Compresses borderline cards to 450 chars so LLM Critic runs 5x faster.
 */
export function partitionCardsForFastCritic(
  cards: CriticCardInput[],
  topic: string
): FastCriticPartition {
  const autoApproved: FastCriticPartition["autoApproved"] = [];
  const autoRejected: FastCriticPartition["autoRejected"] = [];
  const needsLlmGrading: FastCriticPartition["needsLlmGrading"] = [];

  for (const card of cards) {
    const combinedHeader = `${card.title} ${card.url}`;

    // 1. Instant reject dictionary definitions & generic "Search Page" stubs seen in plugin_3.log
    if (
      JUNK_TITLE_OR_URL_REGEX.test(combinedHeader) ||
      card.title.trim().toLowerCase() === "search page" ||
      /^meaning of .+ in english/i.test(card.title.trim())
    ) {
      autoRejected.push({
        id: card.id,
        reason: "heuristic reject: dictionary or search index page",
      });
      continue;
    }

    // 2. Auto-approve Tier-0 verified JSON sources (OpenAlex / Wiki / PubMed)
    const engineLower = (card.engine ?? "").toLowerCase();
    const isCanonicalDomain =
      CANONICAL_ENGINES.has(engineLower) ||
      /\b(pubmed\.ncbi\.nlm\.nih\.gov|ncbi\.nlm\.nih\.gov|doi\.org|arxiv\.org|openalex\.org)\b/i.test(
        card.url
      );

    if (isCanonicalDomain && card.content.length >= 180) {
      autoApproved.push({
        id: card.id,
        tier: engineLower === "openalex" || card.url.includes("pubmed") ? "A" : "B",
        confidence: "High",
        reason: "canonical academic/reference source",
      });
      continue;
    }

    // 3. Auto-approve high lexical relevance cards (rel >= 0.85) with substantive text
    // Upgrade spec: deterministic relevance >= 0.85 AND >= 400 chars of clean prose
    if ((card.relevanceScore ?? 0) >= 0.85 && card.content.length >= 400) {
      autoApproved.push({
        id: card.id,
        tier: "B",
        confidence: "High",
        reason: `high deterministic relevance (${card.relevanceScore?.toFixed(2)})`,
      });
      continue;
    }

    // 4. Send only borderline cards to LLM Critic, with a tight 450-char excerpt
    needsLlmGrading.push({
      ...card,
      compactExcerpt: buildCompactCriticExcerpt(card.content, topic, 450),
    });
  }

  return { autoApproved, autoRejected, needsLlmGrading };
}

function buildCompactCriticExcerpt(
  content: string,
  topic: string,
  maxChars = 450
): string {
  const clean = content.replace(/\s+/g, " ").trim();
  if (clean.length <= maxChars) return clean;

  const terms = topic
    .toLowerCase()
    .split(/[^a-z0-9]+/)
    .filter((t) => t.length > 3);

  const sentences = clean.split(/(?<=[.!?])\s+/);
  const scored = sentences.map((s, idx) => {
    const lower = s.toLowerCase();
    const hits = terms.filter((t) => lower.includes(t)).length;
    return { idx, s, hits };
  });

  scored.sort((a, b) => b.hits - a.hits || a.idx - b.idx);
  let out = "";
  for (const item of scored) {
    if ((out + " " + item.s).trim().length > maxChars) break;
    out = (out + " " + item.s).trim();
  }
  return out || clean.slice(0, maxChars);
}

export interface CriticCardInput {
  id: string;
  url: string;
  title: string;
  content: string;
  engine?: string;
  relevanceScore?: number;
  sourceTier?: string;
}

export interface FastCriticPartition {
  autoApproved: Array<{
    id: string;
    tier: "A" | "B";
    confidence: "High" | "Medium";
    reason: string;
  }>;
  autoRejected: Array<{
    id: string;
    reason: string;
  }>;
  needsLlmGrading: Array<CriticCardInput & { compactExcerpt: string }>;
}

const JUNK_TITLE_OR_URL_REGEX =
  /\b(merriam-webster\.com\/dictionary|dictionary\.cambridge\.org|collinsdictionary\.com|vocabulary\.com|\/search\?|search\s+page|login|sign\s+in)\b/i;

const CANONICAL_ENGINES = new Set([
  "openalex",
  "crossref",
  "arxiv",
  "pubmed",
  "wikipedia",
]);

/**
 * Partitions evidence cards BEFORE calling the LLM Critic:
 * 1. Rejects dictionary/search-page noise ("Search Page", "merriam-webster.com/dictionary/original").
 * 2. Auto-approves Tier-0 canonical academic/reference cards & high-relevance (rel >= 0.85) cards.
 * 3. Compresses borderline cards to 450 chars so LLM Critic runs 5x faster.
 */
export function partitionCardsForFastCritic(
  cards: CriticCardInput[],
  topic: string
): FastCriticPartition {
  const autoApproved: FastCriticPartition["autoApproved"] = [];
  const autoRejected: FastCriticPartition["autoRejected"] = [];
  const needsLlmGrading: FastCriticPartition["needsLlmGrading"] = [];

  for (const card of cards) {
    const combinedHeader = `${card.title} ${card.url}`;

    // 1. Instant reject dictionary definitions & generic "Search Page" stubs seen in plugin_3.log
    if (
      JUNK_TITLE_OR_URL_REGEX.test(combinedHeader) ||
      card.title.trim().toLowerCase() === "search page" ||
      /^meaning of .+ in english/i.test(card.title.trim())
    ) {
      autoRejected.push({
        id: card.id,
        reason: "heuristic reject: dictionary or search index page",
      });
      continue;
    }

    // 2. Auto-approve Tier-0 verified JSON sources (OpenAlex / Wiki / PubMed)
    const engineLower = (card.engine ?? "").toLowerCase();
    const isCanonicalDomain =
      CANONICAL_ENGINES.has(engineLower) ||
      /\b(pubmed\.ncbi\.nlm\.nih\.gov|ncbi\.nlm\.nih\.gov|doi\.org|arxiv\.org|openalex\.org)\b/i.test(
        card.url
      );

    if (isCanonicalDomain && card.content.length >= 180) {
      autoApproved.push({
        id: card.id,
        tier: engineLower === "openalex" || card.url.includes("pubmed") ? "A" : "B",
        confidence: "High",
        reason: "canonical academic/reference source",
      });
      continue;
    }

    // 3. Auto-approve high lexical relevance cards (rel >= 0.85) with substantive text
    // Upgrade spec: deterministic relevance >= 0.85 AND >= 400 chars of clean prose
    if ((card.relevanceScore ?? 0) >= 0.85 && card.content.length >= 400) {
      autoApproved.push({
        id: card.id,
        tier: "B",
        confidence: "High",
        reason: `high deterministic relevance (${card.relevanceScore?.toFixed(2)})`,
      });
      continue;
    }

    // 4. Send only borderline cards to LLM Critic, with a tight 450-char excerpt
    needsLlmGrading.push({
      ...card,
      compactExcerpt: buildCompactCriticExcerpt(card.content, topic, 450),
    });
  }

  return { autoApproved, autoRejected, needsLlmGrading };
}

function buildCompactCriticExcerpt(
  content: string,
  topic: string,
  maxChars = 450
): string {
  const clean = content.replace(/\s+/g, " ").trim();
  if (clean.length <= maxChars) return clean;

  const terms = topic
    .toLowerCase()
    .split(/[^a-z0-9]+/)
    .filter((t) => t.length > 3);

  const sentences = clean.split(/(?<=[.!?])\s+/);
  const scored = sentences.map((s, idx) => {
    const lower = s.toLowerCase();
    const hits = terms.filter((t) => lower.includes(t)).length;
    return { idx, s, hits };
  });

  scored.sort((a, b) => b.hits - a.hits || a.idx - b.idx);
  let out = "";
  for (const item of scored) {
    if ((out + " " + item.s).trim().length > maxChars) break;
    out = (out + " " + item.s).trim();
  }
  return out || clean.slice(0, maxChars);
}