File size: 19,866 Bytes
d82db76
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
#!/usr/bin/env python3
"""Audit page OCR document outputs for page-level evaluation.

This audit is intentionally narrower than `ocrtech audit-limbu-output`: it works
on page-eval-compatible `document.json` directories produced by the page
materializers and filters. It writes `limbu-output-audit.json` into each sample
directory so `tools/run_page_ocr_eval.py --require-output-audit` can enforce the
check before scoring live detector outputs.
"""

from __future__ import annotations

import argparse
import json
import platform
from collections import Counter
from datetime import UTC, datetime
from pathlib import Path
from typing import Any

from ocrtech.manifest import ManifestEntry, load_manifest, sha256_file
from ocrtech.schemas import Document


def _resolve_manifest_path(path_value: str | Path, manifest_dir: Path) -> Path:
    path = Path(path_value)
    if path.is_absolute() or path.exists():
        return path
    candidate = manifest_dir / path
    if candidate.exists():
        return candidate
    return path


def _resolve_existing_path(path_value: object, *, base_dirs: list[Path]) -> Path | None:
    if not isinstance(path_value, str) or not path_value:
        return None
    path = Path(path_value)
    candidates = [path] if path.is_absolute() else [path, *(base / path for base in base_dirs)]
    for candidate in candidates:
        if candidate.is_file():
            return candidate
    return path


def _document_candidates(entry: ManifestEntry, input_path: Path, documents_dir: Path) -> list[Path]:
    return [
        documents_dir / entry.sample_id / "document.json",
        documents_dir / input_path.stem / "document.json",
        documents_dir / f"{entry.sample_id}.json",
        documents_dir / f"{input_path.stem}.json",
    ]


def _load_document(path: Path, issues: list[str]) -> Document | None:
    if not path.is_file():
        issues.append(f"document.json is missing: {path}")
        return None
    try:
        payload = json.loads(path.read_text(encoding="utf-8"))
    except json.JSONDecodeError as exc:
        issues.append(f"document.json is invalid JSON: {path}: {exc}")
        return None
    if not isinstance(payload, dict):
        issues.append(f"document.json must contain an object: {path}")
        return None
    try:
        return Document.from_dict(payload)
    except Exception as exc:
        issues.append(f"document.json failed schema load: {type(exc).__name__}: {exc}")
        return None


def _same_file_or_hash(left: Path, right: Path) -> tuple[bool, str]:
    try:
        if left.resolve() == right.resolve():
            return True, "same_path"
    except OSError:
        if str(left) == str(right):
            return True, "same_path_unresolved"
    if left.is_file() and right.is_file() and sha256_file(left) == sha256_file(right):
        return True, "same_sha256"
    return False, "different"


def _audit_source_binding(
    *,
    sample_id: str,
    document: Document,
    document_dir: Path,
    input_path: Path,
    args: argparse.Namespace,
    issues: list[str],
    warnings: list[str],
) -> dict[str, Any]:
    source_value = document.source_path
    source_path = _resolve_existing_path(source_value, base_dirs=[document_dir, Path.cwd()])
    binding: dict[str, Any] = {
        "document_source_path": source_value,
        "resolved_document_source_path": str(source_path) if source_path is not None else None,
        "eval_input_path": str(input_path),
        "match_kind": "not_checked",
    }
    if not source_value:
        issues.append(f"{sample_id}: document.source_path is missing")
        return binding
    if args.require_source_image and (source_path is None or not source_path.is_file()):
        issues.append(f"{sample_id}: document.source_path is not a readable file: {source_value}")
        return binding
    if source_path is not None and source_path.is_file():
        binding["document_source_sha256"] = sha256_file(source_path)
    if input_path.is_file():
        binding["eval_input_sha256"] = sha256_file(input_path)
    if args.require_document_source_match:
        if source_path is None or not source_path.is_file() or not input_path.is_file():
            issues.append(f"{sample_id}: cannot verify document.source_path against eval input")
            return binding
        matched, match_kind = _same_file_or_hash(source_path, input_path)
        binding["match_kind"] = match_kind
        if not matched:
            issues.append(
                f"{sample_id}: document.source_path does not match eval input by path or SHA-256: "
                f"document={source_value!r} eval={str(input_path)!r}"
            )
    else:
        warnings.append(f"{sample_id}: document.source_path/eval input match was not required")
    return binding


def _audit_page_and_lines(
    *,
    sample_id: str,
    document: Document,
    document_dir: Path,
    args: argparse.Namespace,
    issues: list[str],
    warnings: list[str],
) -> dict[str, Any]:
    line_count = 0
    empty_line_count = 0
    confidence_values: list[float] = []
    structural_roles: Counter[str] = Counter()
    detector_source_passes: Counter[str] = Counter()
    crop_count = 0
    crop_hash_checked = 0
    for page in document.pages:
        if page.width is None or page.width <= 0:
            issues.append(f"{sample_id}: page {page.page_index} width must be positive")
        if page.height is None or page.height <= 0:
            issues.append(f"{sample_id}: page {page.page_index} height must be positive")
        page_line_ids: set[str] = set()
        for index, line in enumerate(page.text_lines):
            line_count += 1
            label = f"{sample_id}: page {page.page_index} line {index} ({line.line_id or 'no-id'})"
            if line.line_id:
                if line.line_id in page_line_ids:
                    issues.append(f"{label}: duplicate line_id on page")
                page_line_ids.add(line.line_id)
            else:
                issues.append(f"{label}: missing line_id")
            if not line.text.strip():
                empty_line_count += 1
                if not args.allow_empty_text:
                    issues.append(f"{label}: empty OCR text")
            if line.confidence is not None:
                confidence_values.append(float(line.confidence))
                if line.confidence < 0 or line.confidence > 1:
                    issues.append(f"{label}: confidence must be between 0 and 1")
            elif args.require_line_confidence:
                issues.append(f"{label}: missing confidence")
            if line.bbox.w <= 0 or line.bbox.h <= 0:
                issues.append(f"{label}: bbox width/height must be positive")
            if page.width and page.height:
                tolerance = float(args.bbox_tolerance_px)
                if line.bbox.x < -tolerance or line.bbox.y < -tolerance:
                    issues.append(f"{label}: bbox starts outside page bounds")
                if line.bbox.right > page.width + tolerance or line.bbox.bottom > page.height + tolerance:
                    issues.append(f"{label}: bbox extends outside page bounds")
            metadata = line.metadata if isinstance(line.metadata, dict) else {}
            role = str(metadata.get("structural_role") or "text")
            structural_roles[role] += 1
            source_pass = metadata.get("detector_source_pass")
            if isinstance(source_pass, str) and source_pass:
                detector_source_passes[source_pass] += 1
            if args.expected_detector_source_pass and source_pass != args.expected_detector_source_pass:
                issues.append(
                    f"{label}: detector_source_pass {source_pass!r} does not match "
                    f"{args.expected_detector_source_pass!r}"
                )
            crop_path = _resolve_existing_path(metadata.get("crop_path"), base_dirs=[document_dir, Path.cwd()])
            if crop_path is not None and isinstance(metadata.get("crop_path"), str):
                crop_count += 1
                if not crop_path.is_file():
                    issues.append(f"{label}: crop_path is not a readable file: {metadata.get('crop_path')}")
                else:
                    recorded_sha = metadata.get("crop_sha256")
                    if isinstance(recorded_sha, str) and recorded_sha:
                        crop_hash_checked += 1
                        actual_sha = sha256_file(crop_path)
                        if actual_sha != recorded_sha:
                            issues.append(
                                f"{label}: crop_sha256 mismatch for {crop_path}: "
                                f"recorded={recorded_sha} actual={actual_sha}"
                            )
            elif args.require_line_crops:
                issues.append(f"{label}: missing crop_path")
        known_line_ids = {line.line_id for line in page.text_lines if line.line_id}
        block_orders: set[int] = set()
        for block in page.blocks:
            if block.order in block_orders:
                issues.append(f"{sample_id}: page {page.page_index} duplicate block order {block.order}")
            block_orders.add(block.order)
            for line_id in block.line_ids:
                if line_id not in known_line_ids:
                    issues.append(f"{sample_id}: block {block.block_id} references missing line_id {line_id!r}")
    if not document.pages:
        issues.append(f"{sample_id}: document has no pages")
    if args.min_line_count is not None and line_count < args.min_line_count:
        issues.append(f"{sample_id}: line_count {line_count} below minimum {args.min_line_count}")
    if args.max_line_count is not None and line_count > args.max_line_count:
        issues.append(f"{sample_id}: line_count {line_count} above maximum {args.max_line_count}")
    metadata = document.metadata if isinstance(document.metadata, dict) else {}
    expected_kind = str(args.expected_ocr_output_kind or "")
    observed_kind = metadata.get("ocr_output_kind")
    if expected_kind and observed_kind != expected_kind:
        issues.append(f"{sample_id}: document metadata ocr_output_kind {observed_kind!r} does not match {expected_kind!r}")
    recorded_line_count = metadata.get("line_count")
    if isinstance(recorded_line_count, int) and recorded_line_count != line_count:
        issues.append(f"{sample_id}: document metadata line_count {recorded_line_count} does not match actual {line_count}")
    elif recorded_line_count is not None and not isinstance(recorded_line_count, int):
        warnings.append(f"{sample_id}: document metadata line_count is not an integer: {recorded_line_count!r}")
    return {
        "page_count": len(document.pages),
        "line_count": line_count,
        "empty_line_count": empty_line_count,
        "average_line_confidence": sum(confidence_values) / len(confidence_values) if confidence_values else None,
        "line_confidence_count": len(confidence_values),
        "structural_role_counts": dict(sorted(structural_roles.items())),
        "detector_source_pass_counts": dict(sorted(detector_source_passes.items())),
        "crop_count": crop_count,
        "crop_hash_checked_count": crop_hash_checked,
    }


def _write_sample_audit(report: dict[str, Any], output_dir: Path) -> None:
    output_dir.mkdir(parents=True, exist_ok=True)
    (output_dir / "limbu-output-audit.json").write_text(
        json.dumps(report, ensure_ascii=False, indent=2) + "\n",
        encoding="utf-8",
    )
    lines = [
        "# Page OCR Document Output Audit",
        "",
        f"- sample: `{report.get('sample_id')}`",
        f"- passed: `{report.get('passed')}`",
        f"- document: `{report.get('document_path')}`",
        f"- eval input: `{report.get('eval_input_path')}`",
        f"- source match: `{report.get('source_binding', {}).get('match_kind')}`",
        f"- line count: `{report.get('line_count')}`",
        f"- policy: `{json.dumps(report.get('policy', {}), ensure_ascii=False, sort_keys=True)}`",
        "",
        "## Issues",
        "",
    ]
    issues = report.get("issues")
    lines.extend(f"- {issue}" for issue in issues) if isinstance(issues, list) and issues else lines.append("- none")
    lines.extend(["", "## Warnings", ""])
    warnings = report.get("warnings")
    lines.extend(f"- {warning}" for warning in warnings) if isinstance(warnings, list) and warnings else lines.append("- none")
    lines.append("")
    (output_dir / "limbu-output-audit.md").write_text("\n".join(lines), encoding="utf-8")


def _audit_sample(entry: ManifestEntry, *, manifest_dir: Path, documents_dir: Path, args: argparse.Namespace) -> dict[str, Any]:
    issues: list[str] = []
    warnings: list[str] = []
    input_path = _resolve_manifest_path(entry.image_path, manifest_dir)
    if not input_path.is_file():
        issues.append(f"{entry.sample_id}: eval input is missing: {input_path}")
    elif entry.sha256 and sha256_file(input_path) != entry.sha256:
        issues.append(f"{entry.sample_id}: eval input SHA-256 mismatch")
    candidates = _document_candidates(entry, input_path, documents_dir)
    document_path = next((candidate for candidate in candidates if candidate.is_file()), candidates[0])
    document = _load_document(document_path, issues)
    source_binding: dict[str, Any] = {}
    summary: dict[str, Any] = {
        "page_count": 0,
        "line_count": 0,
        "empty_line_count": 0,
        "average_line_confidence": None,
        "line_confidence_count": 0,
        "structural_role_counts": {},
        "detector_source_pass_counts": {},
        "crop_count": 0,
        "crop_hash_checked_count": 0,
    }
    if document is not None:
        source_binding = _audit_source_binding(
            sample_id=entry.sample_id,
            document=document,
            document_dir=document_path.parent,
            input_path=input_path,
            args=args,
            issues=issues,
            warnings=warnings,
        )
        summary = _audit_page_and_lines(
            sample_id=entry.sample_id,
            document=document,
            document_dir=document_path.parent,
            args=args,
            issues=issues,
            warnings=warnings,
        )
    policy = {
        "expected_ocr_output_kind": args.expected_ocr_output_kind,
        "expected_detector_source_pass": args.expected_detector_source_pass,
        "require_source_image": args.require_source_image,
        "require_document_source_match": args.require_document_source_match,
        "require_line_crops": args.require_line_crops,
        "require_line_confidence": args.require_line_confidence,
        "allow_empty_text": args.allow_empty_text,
        "bbox_tolerance_px": args.bbox_tolerance_px,
        "min_line_count": args.min_line_count,
        "max_line_count": args.max_line_count,
    }
    report: dict[str, Any] = {
        "created_at": datetime.now(UTC).isoformat(),
        "machine": platform.node(),
        "tool": "tools/audit_page_ocr_documents.py",
        "sample_id": entry.sample_id,
        "eval_input_path": str(input_path),
        "eval_input_sha256": sha256_file(input_path) if input_path.is_file() else None,
        "document_path": str(document_path),
        "document_sha256": sha256_file(document_path) if document_path.is_file() else None,
        "source_binding": source_binding,
        "policy": policy,
        "passed": not issues,
        "issues": issues,
        "warnings": warnings,
        "claim_boundary": "document artifact audit for page OCR outputs; not a text-quality or reference correctness claim",
        **summary,
    }
    _write_sample_audit(report, document_path.parent)
    return report


def audit_page_ocr_documents(args: argparse.Namespace) -> dict[str, Any]:
    manifest = load_manifest(args.eval_manifest)
    if not manifest:
        raise ValueError(f"eval manifest is empty: {args.eval_manifest}")
    documents_dir = Path(args.documents_dir)
    out = Path(args.out) if args.out is not None else documents_dir
    out.mkdir(parents=True, exist_ok=True)
    sample_reports = [
        _audit_sample(entry, manifest_dir=Path(args.eval_manifest).parent, documents_dir=documents_dir, args=args)
        for entry in manifest
    ]
    aggregate: dict[str, Any] = {
        "created_at": datetime.now(UTC).isoformat(),
        "machine": platform.node(),
        "tool": "tools/audit_page_ocr_documents.py",
        "eval_manifest": str(args.eval_manifest),
        "eval_manifest_sha256": sha256_file(Path(args.eval_manifest)),
        "documents_dir": str(documents_dir),
        "sample_count": len(sample_reports),
        "passed_count": sum(1 for report in sample_reports if report.get("passed") is True),
        "failed_count": sum(1 for report in sample_reports if report.get("passed") is not True),
        "line_count": sum(int(report.get("line_count") or 0) for report in sample_reports),
        "empty_line_count": sum(int(report.get("empty_line_count") or 0) for report in sample_reports),
        "passed": all(report.get("passed") is True for report in sample_reports),
        "policy": sample_reports[0].get("policy", {}) if sample_reports else {},
        "samples": sample_reports,
    }
    (out / "page-ocr-documents-audit.json").write_text(
        json.dumps(aggregate, ensure_ascii=False, indent=2) + "\n",
        encoding="utf-8",
    )
    lines = [
        "# Page OCR Documents Audit",
        "",
        f"- passed: `{aggregate['passed']}`",
        f"- eval manifest: `{args.eval_manifest}`",
        f"- documents dir: `{documents_dir}`",
        f"- samples: `{aggregate['sample_count']}`",
        f"- passed samples: `{aggregate['passed_count']}`",
        f"- failed samples: `{aggregate['failed_count']}`",
        f"- lines: `{aggregate['line_count']}`",
        f"- empty lines: `{aggregate['empty_line_count']}`",
        "",
        "## Failed Samples",
        "",
    ]
    failed = [report for report in sample_reports if report.get("passed") is not True]
    if failed:
        for report in failed:
            lines.append(f"- `{report.get('sample_id')}`: {len(report.get('issues') or [])} issues")
    else:
        lines.append("- none")
    lines.append("")
    (out / "page-ocr-documents-audit.md").write_text("\n".join(lines), encoding="utf-8")
    return aggregate


def main(argv: list[str] | None = None) -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--eval-manifest", type=Path, required=True)
    parser.add_argument("--documents-dir", type=Path, required=True)
    parser.add_argument("--out", type=Path)
    parser.add_argument("--expected-ocr-output-kind")
    parser.add_argument("--expected-detector-source-pass")
    parser.add_argument("--no-require-source-image", dest="require_source_image", action="store_false")
    parser.set_defaults(require_source_image=True)
    parser.add_argument("--no-require-document-source-match", dest="require_document_source_match", action="store_false")
    parser.set_defaults(require_document_source_match=True)
    parser.add_argument("--require-line-crops", action="store_true")
    parser.add_argument("--require-line-confidence", action="store_true")
    parser.add_argument("--allow-empty-text", action="store_true")
    parser.add_argument("--bbox-tolerance-px", type=float, default=2.0)
    parser.add_argument("--min-line-count", type=int)
    parser.add_argument("--max-line-count", type=int)
    args = parser.parse_args(argv)
    result = audit_page_ocr_documents(args)
    print(
        f"page-ocr-documents audit {'passed' if result['passed'] else 'failed'} "
        f"samples={result['sample_count']} failed={result['failed_count']}"
    )
    return 0 if result["passed"] else 1


if __name__ == "__main__":
    raise SystemExit(main())