// Markdown-aware chunking: split by headings first, then pack paragraphs into
// size-bounded windows with overlap. Each chunk carries its heading path so the
// embedding (and the citation shown to the user) keeps its structural context.

export interface Chunk {
  text: string; // raw chunk text (without the heading-path prefix)
  headingPath: string[]; // e.g. ["Résumé", "Détails"]
}

export interface ChunkOptions {
  chunkSizeChars: number;
  chunkOverlapChars: number;
}

interface Section {
  headingPath: string[];
  body: string;
}

const HEADING_RE = /^(#{1,6})\s+(.*)$/;

/** Split markdown into sections keyed by their heading path. */
function splitIntoSections(markdown: string): Section[] {
  const lines = markdown.split(/\r?\n/);
  const stack: string[] = []; // stack[level-1] = title
  let current: Section = { headingPath: [], body: "" };
  const sections: Section[] = [];
  let inFence = false;

  const flush = () => {
    if (current.body.trim().length > 0) sections.push(current);
  };

  for (const line of lines) {
    if (/^\s*```/.test(line)) inFence = !inFence;
    const m = inFence ? null : line.match(HEADING_RE);
    if (m) {
      flush();
      const level = m[1].length;
      const title = m[2].trim();
      stack.length = Math.min(stack.length, level - 1);
      stack[level - 1] = title;
      current = { headingPath: stack.slice(0, level), body: "" };
    } else {
      current.body += line + "\n";
    }
  }
  flush();
  return sections;
}

/** Greedily pack paragraphs into windows of ~size chars, with char overlap. */
function packWithOverlap(text: string, size: number, overlap: number): string[] {
  const trimmed = text.trim();
  if (trimmed.length <= size) return trimmed.length ? [trimmed] : [];

  const paragraphs = trimmed.split(/\n{2,}/);
  const out: string[] = [];
  let buf = "";

  const pushBuf = () => {
    const t = buf.trim();
    if (t) out.push(t);
  };

  for (const para of paragraphs) {
    // A single oversized paragraph is hard-split on character windows.
    if (para.length > size) {
      pushBuf();
      buf = "";
      for (let i = 0; i < para.length; i += Math.max(1, size - overlap)) {
        out.push(para.slice(i, i + size));
      }
      continue;
    }
    if (buf.length + para.length + 2 > size && buf.length > 0) {
      pushBuf();
      // start the next buffer with a tail overlap of the previous one
      buf = overlap > 0 ? buf.slice(Math.max(0, buf.length - overlap)) + "\n\n" : "";
    }
    buf += (buf.length ? "\n\n" : "") + para;
  }
  pushBuf();
  return out;
}

export function chunkMarkdown(markdown: string, opts: ChunkOptions): Chunk[] {
  const size = Math.max(200, opts.chunkSizeChars);
  const overlap = Math.max(0, Math.min(opts.chunkOverlapChars, Math.floor(size / 2)));
  const sections = splitIntoSections(markdown);
  const chunks: Chunk[] = [];
  for (const section of sections) {
    for (const piece of packWithOverlap(section.body, size, overlap)) {
      chunks.push({ text: piece, headingPath: section.headingPath });
    }
  }
  return chunks;
}

/** Text actually handed to the embedding model: heading path + chunk body. */
export function embeddingTextFor(pageTitle: string, chunk: Chunk): string {
  const crumbs = [pageTitle, ...chunk.headingPath].filter(Boolean).join(" > ");
  return crumbs ? `${crumbs}\n\n${chunk.text}` : chunk.text;
}

// Markdown-aware chunking: split by headings first, then pack paragraphs into
// size-bounded windows with overlap. Each chunk carries its heading path so the
// embedding (and the citation shown to the user) keeps its structural context.

export interface Chunk {
  text: string; // raw chunk text (without the heading-path prefix)
  headingPath: string[]; // e.g. ["Résumé", "Détails"]
}

export interface ChunkOptions {
  chunkSizeChars: number;
  chunkOverlapChars: number;
}

interface Section {
  headingPath: string[];
  body: string;
}

const HEADING_RE = /^(#{1,6})\s+(.*)$/;

/** Split markdown into sections keyed by their heading path. */
function splitIntoSections(markdown: string): Section[] {
  const lines = markdown.split(/\r?\n/);
  const stack: string[] = []; // stack[level-1] = title
  let current: Section = { headingPath: [], body: "" };
  const sections: Section[] = [];
  let inFence = false;

  const flush = () => {
    if (current.body.trim().length > 0) sections.push(current);
  };

  for (const line of lines) {
    if (/^\s*```/.test(line)) inFence = !inFence;
    const m = inFence ? null : line.match(HEADING_RE);
    if (m) {
      flush();
      const level = m[1].length;
      const title = m[2].trim();
      stack.length = Math.min(stack.length, level - 1);
      stack[level - 1] = title;
      current = { headingPath: stack.slice(0, level), body: "" };
    } else {
      current.body += line + "\n";
    }
  }
  flush();
  return sections;
}

/** Greedily pack paragraphs into windows of ~size chars, with char overlap. */
function packWithOverlap(text: string, size: number, overlap: number): string[] {
  const trimmed = text.trim();
  if (trimmed.length <= size) return trimmed.length ? [trimmed] : [];

  const paragraphs = trimmed.split(/\n{2,}/);
  const out: string[] = [];
  let buf = "";

  const pushBuf = () => {
    const t = buf.trim();
    if (t) out.push(t);
  };

  for (const para of paragraphs) {
    // A single oversized paragraph is hard-split on character windows.
    if (para.length > size) {
      pushBuf();
      buf = "";
      for (let i = 0; i < para.length; i += Math.max(1, size - overlap)) {
        out.push(para.slice(i, i + size));
      }
      continue;
    }
    if (buf.length + para.length + 2 > size && buf.length > 0) {
      pushBuf();
      // start the next buffer with a tail overlap of the previous one
      buf = overlap > 0 ? buf.slice(Math.max(0, buf.length - overlap)) + "\n\n" : "";
    }
    buf += (buf.length ? "\n\n" : "") + para;
  }
  pushBuf();
  return out;
}

export function chunkMarkdown(markdown: string, opts: ChunkOptions): Chunk[] {
  const size = Math.max(200, opts.chunkSizeChars);
  const overlap = Math.max(0, Math.min(opts.chunkOverlapChars, Math.floor(size / 2)));
  const sections = splitIntoSections(markdown);
  const chunks: Chunk[] = [];
  for (const section of sections) {
    for (const piece of packWithOverlap(section.body, size, overlap)) {
      chunks.push({ text: piece, headingPath: section.headingPath });
    }
  }
  return chunks;
}

/** Text actually handed to the embedding model: heading path + chunk body. */
export function embeddingTextFor(pageTitle: string, chunk: Chunk): string {
  const crumbs = [pageTitle, ...chunk.headingPath].filter(Boolean).join(" > ");
  return crumbs ? `${crumbs}\n\n${chunk.text}` : chunk.text;
}