Momento/memento-note/lib/ai/services/slide-source-assets.ts

/**
 * Server-side extraction of usable assets from a note BEFORE any slide LLM call.
 * Inspired by PPTAgent (retrieve relevant content per slide) and DeckForge (data_needs).
 * Critical for STEM notes: formulas must be harvested, not hoped for from the model.
 */

import { extractPublishImageUrls } from '@/lib/publish/process-note-html'

export interface SourceAssets {
  formulas: string[]
  numbers: Array<{ label: string; value: number; raw: string }>
  keySentences: string[]
  hasMath: boolean
  hasNumbers: boolean
  hasImages: boolean
  wordCount: number
  images: string[]
}

/** Extract LaTeX / equation-like fragments from note plain text or HTML. */
export function extractFormulas(raw: string): string[] {
  if (!raw) return []
  const found: string[] = []
  // 3+ consecutive plain words at brace depth 0 = prose tail captured after a
  // formula (KaTeX eats spaces in math mode → must never reach $…$).
  const PROSE_TAIL_RE =
    /(?:[.!?]?\s+)[A-ZÀ-ÖØ-Þ]?[a-zA-ZÀ-ÿ'’]{2,}\s+[a-zA-ZÀ-ÿ'’]{2,}\s+[a-zA-ZÀ-ÿ'’]{2,}[\s\S]*$/g
  const braceDepth = (s: string): number => {
    let d = 0
    for (const ch of s) {
      if (ch === '{') d++
      else if (ch === '}') d = Math.max(0, d - 1)
    }
    return d
  }
  const push = (s: string) => {
    let t = s.replace(/\s+/g, ' ').trim()
    t = t.replace(/^\$+|\$+$/g, '').trim()
    PROSE_TAIL_RE.lastIndex = 0
    let m: RegExpExecArray | null
    while ((m = PROSE_TAIL_RE.exec(t)) !== null) {
      if (braceDepth(t.slice(0, m.index)) === 0) {
        t = t.slice(0, m.index).trim()
        break
      }
    }
    t = t.replace(/^\$+|\$+$/g, '').trim()
    if (t.length >= 2 && t.length <= 280 && !found.includes(t)) found.push(t)
  }

  // TipTap / published HTML: data-latex="..."
  for (const m of raw.matchAll(/data-latex=["']([^"']+)["']/gi)) {
    push(decodeHtmlEntities(m[1] || ''))
  }

  // $$ ... $$
  for (const m of raw.matchAll(/\$\$([\s\S]+?)\$\$/g)) push(m[1] || '')
  // \[ ... \]
  for (const m of raw.matchAll(/\\\[([\s\S]+?)\\\]/g)) push(m[1] || '')
  // \( ... \)
  for (const m of raw.matchAll(/\\\(([\s\S]+?)\\\)/g)) push(m[1] || '')
  // $ ... $ (single line)
  for (const m of raw.matchAll(/(?<![\\$])\$([^$\n]{2,120})\$(?!\$)/g)) push(m[1] || '')
  // Common latex commands in free text
  for (const m of raw.matchAll(
    /(?:^|[\s(])((?:\\frac\{[^}]+\}\{[^}]+\}|\\partial|\\nabla|\\sum|\\int|\\frac|\\cdot|\\times|\\left|\\right|y'|y''|d[xy]\/d[xy]|∂)[^\n]{0,100})/g,
  )) {
    push(m[1] || '')
  }
  // Equation-like: f(x) = ..., dy/dx = ...
  for (const m of raw.matchAll(
    /(?:^|\n)\s*([A-Za-zΑ-ω∂∇][A-Za-z0-9_'"′]*\s*(?:\([^)]*\))?\s*[=≈≡]\s*[^\n]{3,100})/g,
  )) {
    const line = (m[1] || '').trim()
    if (/[=≈]/.test(line) && /[a-zA-Z0-9]/.test(line)) push(line)
  }

  return found.slice(0, 24)
}

function decodeHtmlEntities(s: string): string {
  return s
    .replace(/&quot;/g, '"')
    .replace(/&#39;/g, "'")
    .replace(/&lt;/g, '<')
    .replace(/&gt;/g, '>')
    .replace(/&amp;/g, '&')
}

export function extractNumbers(raw: string): Array<{ label: string; value: number; raw: string }> {
  if (!raw) return []
  // Normalize Persian (۰-۹) and Arabic (٠-٩) numerals to Western (0-9)
  const text = raw
    .replace(/[\u06F0-\u06F9]/g, (c) => String.fromCharCode(c.charCodeAt(0) - 0x06f0 + 0x30))
    .replace(/[\u0660-\u0669]/g, (c) => String.fromCharCode(c.charCodeAt(0) - 0x0660 + 0x30))

  const out: Array<{ label: string; value: number; raw: string }> = []
  const re =
    /(?:^|[^\d])((?:≈|~)?\s*-?\d+(?:[.,]\d+)?)\s*(%|€|\$|k|m|bn|ms|s|°C)?\b/gi
  let m: RegExpExecArray | null
  while ((m = re.exec(text)) !== null && out.length < 20) {
    const numStr = (m[1] || '').replace(/[≈~\s]/g, '').replace(',', '.')
    const value = parseFloat(numStr)
    if (!Number.isFinite(value)) continue
    const unit = m[2] || ''
    // crude label: 40 chars before
    const start = Math.max(0, m.index - 40)
    const ctx = text.slice(start, m.index).replace(/\s+/g, ' ').trim()
    const label = ctx.split(/[.;:!?\n]/).pop()?.trim().slice(-32) || `n${out.length + 1}`
    out.push({ label, value: unit === '%' ? value : value, raw: `${numStr}${unit}` })
  }
  return out
}

export function extractKeySentences(raw: string, max = 16): string[] {
  if (!raw) return []
  const plain = raw
    .replace(/<[^>]+>/g, ' ')
    .replace(/\$\$[\s\S]+?\$\$/g, ' ')
    .replace(/\\\[[\s\S]+?\\\]/g, ' ')
    .replace(/\s+/g, ' ')
    .trim()
  const parts = plain
    .split(/(?<=[.!?])\s+/)
    .map((s) => s.trim())
    .filter((s) => s.length >= 40 && s.length <= 220)
  const out: string[] = []
  for (const p of parts) {
    if (out.length >= max) break
    if (!out.some((x) => x.slice(0, 40) === p.slice(0, 40))) out.push(p)
  }
  return out
}

export function extractSourceAssets(raw: string): SourceAssets {
  const plain = raw.replace(/<[^>]+>/g, ' ')
  const formulas = extractFormulas(plain)
  const numbers = extractNumbers(plain)
  const keySentences = extractKeySentences(raw)
  const images = extractPublishImageUrls(raw).slice(0, 8)
  const wordCount = plain.split(/\s+/).filter(Boolean).length
  return {
    formulas,
    numbers,
    keySentences,
    hasMath: formulas.length > 0 || /équat|equat|différen|differen|dériv|deriv|intégr|integr|EDO|ODE|PDE|latex/i.test(plain),
    hasNumbers: numbers.length >= 2,
    hasImages: images.length > 0,
    wordCount,
    images,
  }
}