pn-new-crm/src/lib/services/berth-pdf-parser.ts

/**
 * Reverse parser for per-berth PDFs (Phase 6b — see plan §4.7b and §9.2).
 *
 * Three tiers, each falling back to the next:
 *
 *   1. AcroForm — read named text fields via pdf-lib. The sample
 *      `Berth_Spec_Sheet_A1.pdf` has 0 AcroForm fields (designers export the
 *      PDF flat), so this tier is built defensively for future templates that
 *      may include named form fields. When fields exist, this is the highest-
 *      confidence path because there's no OCR loss.
 *
 *   2. OCR — Tesseract.js extracts text from the page; positional/regex
 *      heuristics keyed off the labels documented in §9.2 pull out values.
 *      Returns per-field confidence scores.
 *
 *   3. AI fallback — gated on `getResolvedOcrConfig(...)` returning a usable
 *      OpenAI/Claude config. Only invoked when OCR confidence is below
 *      threshold for too many fields AND the rep opts in via the diff dialog.
 *      A null `apiKey` causes this tier to return a clear "not configured"
 *      error rather than silently falling back to OCR-only.
 */

import { PDFDocument } from 'pdf-lib';

// ─── shared types ────────────────────────────────────────────────────────────

export type ParserEngine = 'acroform' | 'ocr' | 'ai';

/**
 * Canonical extracted shape. Keys map 1:1 to nullable columns on the `berths`
 * table; `mooringNumber` is special (used for the §14.6 mismatch warning).
 */
export interface ExtractedBerthFields {
  mooringNumber?: string | null;
  lengthFt?: number | null;
  lengthM?: number | null;
  widthFt?: number | null;
  widthM?: number | null;
  /** Water depth at the berth (separate from a vessel's max draft). */
  waterDepth?: number | null;
  waterDepthM?: number | null;
  /** Max draught of vessel — falls back to the berth's draft column. */
  draftFt?: number | null;
  draftM?: number | null;
  bowFacing?: string | null;
  sidePontoon?: string | null;
  powerCapacity?: number | null;
  voltage?: number | null;
  mooringType?: string | null;
  cleatType?: string | null;
  cleatCapacity?: string | null;
  bollardType?: string | null;
  bollardCapacity?: string | null;
  access?: string | null;
  weeklyRateHighUsd?: number | null;
  weeklyRateLowUsd?: number | null;
  dailyRateHighUsd?: number | null;
  dailyRateLowUsd?: number | null;
  /** ISO date YYYY-MM-DD. */
  pricingValidUntil?: string | null;
  price?: number | null;
}

export interface ParsedField<T = unknown> {
  value: T;
  /** 0..1 confidence; 1 means "absolute match" (AcroForm or unambiguous regex). */
  confidence: number;
  /** Engine that produced this field; helps the diff dialog explain itself. */
  engine: ParserEngine;
}

export interface ParseResult {
  engine: ParserEngine;
  /** Sparse — only fields the parser was able to extract. */
  fields: Partial<Record<keyof ExtractedBerthFields, ParsedField>>;
  /** Mean confidence across all extracted fields (0..1). */
  meanConfidence: number;
  /** Raw text the OCR or AI tier produced — useful for the diff dialog audit. */
  rawText?: string;
  /** Set when a tier degraded; the API surface uses this to decide whether to
   *  surface the "AI parse" button. */
  warnings: string[];
}

// ─── magic-byte check (§14.6 critical) ───────────────────────────────────────

/** Reads first 5 bytes; returns true iff they are `%PDF-`. */
export function isPdfMagic(buffer: Buffer): boolean {
  if (buffer.length < 5) return false;
  return (
    buffer[0] === 0x25 && // %
    buffer[1] === 0x50 && // P
    buffer[2] === 0x44 && // D
    buffer[3] === 0x46 && // F
    buffer[4] === 0x2d //   -
  );
}

// ─── tier 1: AcroForm ────────────────────────────────────────────────────────

/**
 * AcroForm field name → ExtractedBerthFields key. Mirrors the names §4.7b
 * mentions ("length_ft", "mooring_number"…) plus a couple of tolerant aliases.
 */
const ACROFORM_FIELD_MAP: Record<string, keyof ExtractedBerthFields> = {
  mooring_number: 'mooringNumber',
  berth_number: 'mooringNumber',
  length_ft: 'lengthFt',
  length_m: 'lengthM',
  width_ft: 'widthFt',
  width_m: 'widthM',
  draft_ft: 'draftFt',
  draft_m: 'draftM',
  water_depth: 'waterDepth',
  water_depth_m: 'waterDepthM',
  bow_facing: 'bowFacing',
  side_pontoon: 'sidePontoon',
  pontoon: 'sidePontoon',
  power_capacity: 'powerCapacity',
  voltage: 'voltage',
  mooring_type: 'mooringType',
  cleat_type: 'cleatType',
  cleat_capacity: 'cleatCapacity',
  bollard_type: 'bollardType',
  bollard_capacity: 'bollardCapacity',
  access: 'access',
  weekly_rate_high_usd: 'weeklyRateHighUsd',
  weekly_rate_low_usd: 'weeklyRateLowUsd',
  daily_rate_high_usd: 'dailyRateHighUsd',
  daily_rate_low_usd: 'dailyRateLowUsd',
  pricing_valid_until: 'pricingValidUntil',
  price: 'price',
};

async function tryAcroForm(buffer: Buffer): Promise<ParseResult | null> {
  let doc: PDFDocument;
  try {
    doc = await PDFDocument.load(buffer, { ignoreEncryption: true });
  } catch {
    return null;
  }
  let form: ReturnType<PDFDocument['getForm']>;
  try {
    form = doc.getForm();
  } catch {
    return null;
  }
  const fields = form.getFields();
  if (fields.length === 0) return null;

  const out: Partial<Record<keyof ExtractedBerthFields, ParsedField>> = {};
  for (const field of fields) {
    const name = field.getName().toLowerCase();
    const target = ACROFORM_FIELD_MAP[name];
    if (!target) continue;
    // pdf-lib doesn't expose a generic "get value" — narrow to text fields.
    let raw: string | undefined;
    try {
      const tf = form.getTextField(field.getName());
      raw = tf.getText() ?? undefined;
    } catch {
      continue;
    }
    if (!raw || raw.trim().length === 0) continue;
    const parsed = coerceFieldValue(target, raw.trim());
    if (parsed === null) continue;
    out[target] = { value: parsed, confidence: 1, engine: 'acroform' };
  }

  if (Object.keys(out).length === 0) return null;
  return {
    engine: 'acroform',
    fields: out,
    meanConfidence: 1,
    warnings: [],
  };
}

// ─── tier 2: OCR via Tesseract ───────────────────────────────────────────────

/**
 * Runs Tesseract against a PDF rasterized to one image per page. Tesseract.js
 * accepts image inputs; we use a lazy `pdfjs-dist`-style rasterization fallback
 * via dynamic import. To keep the parser unit-testable without a WASM bundle,
 * the actual recognize() call is encapsulated in the `runOcr` adapter that
 * production wires to tesseract.js and tests can stub.
 */
export interface OcrAdapter {
  /** Returns plain text + a 0..100 mean confidence score. */
  recognize(buffer: Buffer): Promise<{ text: string; confidence: number }>;
}

/** Hard cap on Tesseract OCR runtime. A crafted PDF rasterizing to
 *  high-resolution noise can pin the process indefinitely (CPU bomb).
 *  30 seconds covers the legitimate single-page-spec case by a wide
 *  margin while bounding the worst-case worker hold-time. The AI
 *  fallback tier handles cases where OCR couldn't finish. */
const OCR_TIMEOUT_MS = 30_000;

/** Default adapter — dynamically imports tesseract.js so the WASM bundle isn't
 *  pulled into client builds. */
async function defaultOcrAdapter(): Promise<OcrAdapter> {
  return {
    recognize: async (buffer: Buffer) => {
      const tesseract = await import('tesseract.js');
      // Race the OCR against a timeout so a runaway recognition can't
      // hold the worker forever. The race-loser pattern doesn't
      // actually cancel Tesseract (no AbortController support), but it
      // does free the awaiter so the caller can fall through to AI.
      let timeoutHandle: NodeJS.Timeout | undefined;
      const timeout = new Promise<{ text: string; confidence: number }>((_, reject) => {
        timeoutHandle = setTimeout(
          () => reject(new Error(`Tesseract OCR exceeded ${OCR_TIMEOUT_MS}ms timeout`)),
          OCR_TIMEOUT_MS,
        );
      });
      try {
        const result = await Promise.race([
          tesseract.recognize(buffer, 'eng').then((r) => ({
            text: r.data.text ?? '',
            confidence: typeof r.data.confidence === 'number' ? r.data.confidence : 0,
          })),
          timeout,
        ]);
        return result;
      } finally {
        if (timeoutHandle) clearTimeout(timeoutHandle);
      }
    },
  };
}

/**
 * Heuristic extraction from OCR text. The patterns mirror the layout
 * documented in plan §9.2:
 *
 *   - "Length: 206' 8" / 63m"
 *   - "Mooring: A12" or large "A1" near "BERTH NUMBER"
 *   - "WEEK HIGH / LOW" and "DAY HIGH / LOW" pricing blocks
 *   - "ALL PRICES ABOVE ARE CONFIRMED THROUGH UNTIL <date>"
 */
export function extractFromOcrText(rawText: string): {
  fields: Partial<Record<keyof ExtractedBerthFields, ParsedField>>;
  warnings: string[];
} {
  const warnings: string[] = [];
  const out: Partial<Record<keyof ExtractedBerthFields, ParsedField>> = {};

  // Normalize whitespace for line-based regexes but keep structure.
  const text = rawText.replace(/ /g, ' ');

  // Mooring number: BERTH NUMBER block. We try a couple of layouts.
  const mooringMatch =
    text.match(/BERTH\s+NUMBER[\s\S]{0,80}?\b([A-Z]\d{1,3})\b/i) ??
    text.match(/^\s*([A-Z]\d{1,3})\s*$/m) ??
    text.match(/Mooring(?:\s+Number)?\s*[:#]?\s*([A-Z]\d{1,3})/i);
  if (mooringMatch) {
    out.mooringNumber = { value: mooringMatch[1]!.toUpperCase(), confidence: 0.85, engine: 'ocr' };
  }

  // Length / Width / Water Depth — `Label: <imperial> / <metric>` form.
  // Imperial may be `206' 8"` style; we capture the numeric prefix in feet
  // and parse the metric independently because they're rarely lossless.
  const dimensional = (
    label: string,
    ftKey: keyof ExtractedBerthFields,
    mKey: keyof ExtractedBerthFields,
  ) => {
    const re = new RegExp(
      `${label}\\s*[:.]?\\s*([0-9]+(?:'\\s*[0-9]+\")?(?:\\.[0-9]+)?)\\s*(?:ft)?\\s*\\/\\s*([0-9]+(?:\\.[0-9]+)?)\\s*m`,
      'i',
    );
    const m = text.match(re);
    if (!m) return;
    const ft = parseFeetInches(m[1]!);
    const meters = Number(m[2]);
    if (ft != null && Number.isFinite(ft)) {
      out[ftKey] = { value: ft, confidence: 0.8, engine: 'ocr' } as ParsedField;
    }
    if (Number.isFinite(meters)) {
      out[mKey] = { value: meters, confidence: 0.85, engine: 'ocr' } as ParsedField;
    }
    if (ft != null && Number.isFinite(meters) && Math.abs(ft * 0.3048 - meters) / meters > 0.01) {
      warnings.push(
        `${label}: imperial/metric mismatch — ${ft}ft vs ${meters}m differ >1% (using imperial as source of truth).`,
      );
    }
  };
  dimensional('Length', 'lengthFt', 'lengthM');
  dimensional('Width', 'widthFt', 'widthM');
  dimensional('Water\\s+Depth', 'waterDepth', 'waterDepthM');
  // Max draft of vessel maps to the berth's draft column.
  dimensional('Max\\.?\\s*draught(?:\\s+of\\s+vessel)?', 'draftFt', 'draftM');

  // Singular labels (`Bow Facing: East`, `Pontoon: QUAY PT`).
  const labelToKey: Array<[RegExp, keyof ExtractedBerthFields]> = [
    [/Bow\s+Facing\s*[:.]?\s*([A-Za-z .]+?)(?:\n|$)/i, 'bowFacing'],
    [/Pontoon\s*[:.]?\s*([A-Za-z0-9 .\-]+?)(?:\n|$)/i, 'sidePontoon'],
    [/Mooring\s+Type\s*[:.]?\s*([A-Za-z0-9 \-\/]+?)(?:\n|$)/i, 'mooringType'],
    [/Cleat\s+Type\s*[:.]?\s*([A-Za-z0-9 \-]+?)(?:\n|$)/i, 'cleatType'],
    [/Cleat\s+Capacity\s*[:.]?\s*([A-Za-z0-9 \-]+?)(?:\n|$)/i, 'cleatCapacity'],
    [/Bollard\s+Type\s*[:.]?\s*([A-Za-z0-9 \-]+?)(?:\n|$)/i, 'bollardType'],
    [/Bollard\s+Capacity\s*[:.]?\s*([A-Za-z0-9 \-]+?)(?:\n|$)/i, 'bollardCapacity'],
    [/Access\s*[:.]?\s*([A-Za-z0-9 .,()\-]+?)(?:\n|$)/i, 'access'],
  ];
  for (const [re, key] of labelToKey) {
    const m = text.match(re);
    if (m && m[1]) {
      out[key] = { value: m[1].trim(), confidence: 0.75, engine: 'ocr' } as ParsedField;
    }
  }

  // Power Capacity (kW) and Voltage at 60Hz.
  const powerMatch = text.match(/Power\s+Capacity\s*[:.]?\s*([0-9]+(?:\.[0-9]+)?)\s*kW/i);
  if (powerMatch) {
    out.powerCapacity = { value: Number(powerMatch[1]), confidence: 0.85, engine: 'ocr' };
  }
  const voltageMatch = text.match(/Voltage(?:\s+at\s+60\s*Hz)?\s*[:.]?\s*([0-9]+)\s*V/i);
  if (voltageMatch) {
    out.voltage = { value: Number(voltageMatch[1]), confidence: 0.85, engine: 'ocr' };
  }

  // Pricing: "WEEK HIGH / LOW: 11,341 USD / 8,100 USD"
  const weekMatch = text.match(
    /WEEK\s+HIGH\s*\/\s*LOW[:.\s]*([0-9,]+)\s*USD\s*\/\s*([0-9,]+)\s*USD/i,
  );
  if (weekMatch) {
    out.weeklyRateHighUsd = {
      value: Number(weekMatch[1]!.replace(/,/g, '')),
      confidence: 0.8,
      engine: 'ocr',
    };
    out.weeklyRateLowUsd = {
      value: Number(weekMatch[2]!.replace(/,/g, '')),
      confidence: 0.8,
      engine: 'ocr',
    };
  }
  const dayMatch = text.match(
    /DAY\s+HIGH\s*\/\s*LOW[:.\s]*([0-9,]+)\s*USD\s*\/\s*([0-9,]+)\s*USD/i,
  );
  if (dayMatch) {
    out.dailyRateHighUsd = {
      value: Number(dayMatch[1]!.replace(/,/g, '')),
      confidence: 0.8,
      engine: 'ocr',
    };
    out.dailyRateLowUsd = {
      value: Number(dayMatch[2]!.replace(/,/g, '')),
      confidence: 0.8,
      engine: 'ocr',
    };
  }

  // Purchase price: "PURCHASE PRICE:\nFEE SIMPLE OR STRATA LOT\n3,880,800 USD"
  const priceMatch = text.match(/PURCHASE\s+PRICE[\s\S]{0,80}?([0-9][0-9,]+)\s*USD/i);
  if (priceMatch) {
    out.price = { value: Number(priceMatch[1]!.replace(/,/g, '')), confidence: 0.7, engine: 'ocr' };
  }

  // Pricing validity: "ALL PRICES ABOVE ARE CONFIRMED THROUGH UNTIL SEPTEMBER 15TH, 2025"
  const validityMatch = text.match(
    /CONFIRMED\s+THROUGH\s+UNTIL\s+([A-Za-z]+\s+[0-9]{1,2})(?:[A-Z]{2})?,?\s+([0-9]{4})/i,
  );
  if (validityMatch) {
    const iso = parseHumanDate(`${validityMatch[1]} ${validityMatch[2]}`);
    if (iso) {
      out.pricingValidUntil = { value: iso, confidence: 0.75, engine: 'ocr' };
    } else {
      warnings.push(
        'Could not normalize "CONFIRMED THROUGH UNTIL" date; pricing_valid_until skipped.',
      );
    }
  }

  return { fields: out, warnings };
}

async function tryOcr(buffer: Buffer, adapter?: OcrAdapter): Promise<ParseResult | null> {
  const ocr = adapter ?? (await defaultOcrAdapter());
  const result = await ocr.recognize(buffer);
  if (!result.text || result.text.length === 0) {
    return {
      engine: 'ocr',
      fields: {},
      meanConfidence: 0,
      rawText: '',
      warnings: ['OCR produced no text.'],
    };
  }
  const { fields, warnings } = extractFromOcrText(result.text);
  // Tesseract gives 0..100; normalize to 0..1 and use it as a global floor —
  // per-field confidence is set by the regex tier above.
  const floor = Math.max(0, Math.min(result.confidence, 100)) / 100;
  for (const key of Object.keys(fields) as Array<keyof ExtractedBerthFields>) {
    const f = fields[key];
    if (f) f.confidence = Math.min(f.confidence, Math.max(floor, 0.5));
  }
  const values = Object.values(fields);
  const meanConfidence =
    values.length === 0
      ? 0
      : values.reduce((sum, v) => sum + (v?.confidence ?? 0), 0) / values.length;
  return {
    engine: 'ocr',
    fields,
    meanConfidence,
    rawText: result.text,
    warnings,
  };
}

// ─── tier 3: AI fallback ─────────────────────────────────────────────────────

/** Confidence floor below which we recommend the AI tier in the diff dialog. */
export const OCR_LOW_CONFIDENCE_THRESHOLD = 0.55;

/** True when the rep should be offered an "AI parse" button. */
export function shouldOfferAiTier(parse: ParseResult): boolean {
  if (parse.engine !== 'ocr') return false;
  if (Object.keys(parse.fields).length === 0) return true;
  return parse.meanConfidence < OCR_LOW_CONFIDENCE_THRESHOLD;
}

// ─── public entry point ──────────────────────────────────────────────────────

export interface ParseBerthPdfOptions {
  /** Override Tesseract for testing. Production flows resolve the default. */
  ocrAdapter?: OcrAdapter;
  /** Skip the OCR tier when only AcroForm is wanted (e.g. unit tests). */
  skipOcr?: boolean;
}

/**
 * Parse a per-berth PDF buffer. Each tier falls back to the next; the
 * returned result's `engine` field tells callers which tier produced the
 * fields (used by the reconcile-diff dialog to colour confidence chips).
 *
 * The AI tier is never invoked from this entry point — that's a separate
 * deliberate action triggered from the diff dialog so OPENAI_API_KEY isn't
 * spent on every upload.
 */
export async function parseBerthPdf(
  buffer: Buffer,
  opts: ParseBerthPdfOptions = {},
): Promise<ParseResult> {
  if (!isPdfMagic(buffer)) {
    throw new Error('PDF magic-byte check failed: file does not begin with %PDF-');
  }
  const acro = await tryAcroForm(buffer);
  if (acro && Object.keys(acro.fields).length > 0) return acro;
  if (opts.skipOcr) {
    return {
      engine: 'ocr',
      fields: {},
      meanConfidence: 0,
      warnings: ['skipOcr=true; no AcroForm fields found.'],
    };
  }
  const ocr = await tryOcr(buffer, opts.ocrAdapter);
  return (
    ocr ?? {
      engine: 'ocr',
      fields: {},
      meanConfidence: 0,
      warnings: ['OCR adapter returned null.'],
    }
  );
}

// ─── helpers ─────────────────────────────────────────────────────────────────

/** Coerce an AcroForm raw value to the right scalar for the target column. */
function coerceFieldValue(key: keyof ExtractedBerthFields, raw: string): string | number | null {
  // String columns
  const stringKeys: Array<keyof ExtractedBerthFields> = [
    'mooringNumber',
    'bowFacing',
    'sidePontoon',
    'mooringType',
    'cleatType',
    'cleatCapacity',
    'bollardType',
    'bollardCapacity',
    'access',
    'pricingValidUntil',
  ];
  if (stringKeys.includes(key)) {
    if (key === 'pricingValidUntil') {
      // Accept ISO YYYY-MM-DD as-is; otherwise try a humane parse.
      if (/^\d{4}-\d{2}-\d{2}$/.test(raw)) return raw;
      return parseHumanDate(raw);
    }
    return raw;
  }
  // Numeric columns: strip currency / unit suffixes and commas. Berth
  // dimensions / capacities / prices are all non-negative — reject
  // negatives outright so an AcroForm with `length_ft="-50"` doesn't
  // poison the recommender feasibility filter when applied.
  const numeric = Number(raw.replace(/[^0-9.\-]/g, ''));
  if (!Number.isFinite(numeric)) return null;
  return numeric < 0 ? null : numeric;
}

/** Parse a human date like "September 15 2025" → "2025-09-15". */
export function parseHumanDate(raw: string): string | null {
  const cleaned = raw.replace(/(\d+)(st|nd|rd|th)/i, '$1').trim();
  // Force UTC interpretation by appending a Z; otherwise dates without an
  // explicit zone get parsed in the runner's local TZ and `toISOString()`
  // shifts the day by ±1 (caught a -0700 -> 09-14 regression locally).
  const d = new Date(cleaned + ' UTC');
  if (Number.isNaN(d.getTime())) return null;
  return d.toISOString().slice(0, 10);
}

/** Convert "206' 8\"" or "82" → 206.667 / 82. Returns null on parse failure. */
export function parseFeetInches(raw: string): number | null {
  const trimmed = raw.trim();
  const ftIn = trimmed.match(/^([0-9]+)\s*'\s*([0-9]+)\s*"$/);
  if (ftIn) {
    return Number(ftIn[1]) + Number(ftIn[2]) / 12;
  }
  const ftOnly = trimmed.match(/^([0-9]+(?:\.[0-9]+)?)/);
  if (ftOnly) return Number(ftOnly[1]);
  return null;
}