File size: 23,626 Bytes
e71fabd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58259d1
e71fabd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58259d1
e71fabd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58259d1
e71fabd
 
 
 
 
 
 
58259d1
e71fabd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
import os
import yaml
from typing import List, Dict, Any, Optional
from pathlib import Path
import PyPDF2
from .utils import Chunk, TextProcessor, generate_id
import logging as _logging
_logger = _logging.getLogger("rag_ingest")
import os as _os
_OPENAI_ENABLED = False
try:
    from openai import OpenAI as _OpenAI
    _OPENAI_ENABLED = True if _os.getenv("OPENAI_API_KEY") else False
except Exception:
    _OPENAI_ENABLED = False

class OpenAIMetadataDetector:
    """Use OpenAI to detect language, doc_type, and hierarchy levels for a chunk.
    Falls back to heuristics when OpenAI is not available.
    """
    def __init__(self, hierarchy_manager: 'HierarchyManager'):
        self.hierarchy_manager = hierarchy_manager
        self.client = _OpenAI() if _OPENAI_ENABLED else None
        self.model = _os.getenv("OPENAI_MODEL", "gpt-4o-mini")

    def detect(self, text: str) -> Dict[str, Any]:
        if not self.client:
            return {}
        hierarchies = self.hierarchy_manager.list_hierarchies()
        prompt = (
            "You are a metadata extractor. Given a text chunk, infer: language (en|ja), "
            "document_type (Policy|Manual|FAQ|Report|Note|Guideline), hierarchy_name, level1, level2, level3. "
            "CRITICAL: hierarchy_name MUST be exactly one of the following: "
            f"{hierarchies}. Do not invent other names. "
            "Respond as strict JSON with keys: language, document_type, hierarchy_name, level1, level2, level3. "
            "Be concise; if unsure, pick the closest.\n\nText:\n" + text[:2000]
        )
        try:
            _logger.debug("Calling OpenAI for chunk metadata detection (model=%s)", self.model)
            resp = self.client.chat.completions.create(
                model=self.model,
                messages=[{"role": "user", "content": prompt}],
                temperature=0.0,
            )
            content = resp.choices[0].message.content
            import json as _json
            data = _json.loads(content)
            # Enforce allowed hierarchy set
            if isinstance(data, dict) and data.get("hierarchy_name") not in hierarchies:
                data["hierarchy_name"] = None
            _logger.debug("OpenAI chunk metadata inferred: %s", data)
            return data if isinstance(data, dict) else {}
        except Exception:
            _logger.exception("OpenAI chunk metadata detection failed; using heuristics.")
            return {}

# Try to import pypdf (newer, more robust PDF library)
try:
    from pypdf import PdfReader as PyPdfReader
    PYPDF_AVAILABLE = True
except ImportError:
    PYPDF_AVAILABLE = False

class DocumentLoader:
    """Load documents from various formats"""
    
    def __init__(self):
        self.text_processor = TextProcessor()
    
    def load_pdf(self, file_path: str) -> str:
        """Load text from PDF file with fallback readers, preserving paragraphs"""
        # Validate file exists and is readable
        if not os.path.exists(file_path):
            raise FileNotFoundError(f"PDF file not found: {file_path}")
        
        if not os.path.isfile(file_path):
            raise ValueError(f"Path is not a file: {file_path}")
        
        # Check file size
        file_size = os.path.getsize(file_path)
        if file_size == 0:
            raise ValueError(f"PDF file is empty: {file_path}")
        
        # Try pypdf first (more robust)
        if PYPDF_AVAILABLE:
            try:
                with open(file_path, 'rb') as file:
                    reader = PyPdfReader(file)
                    text = ""
                    for page in reader.pages:
                        page_text = page.extract_text()
                        if page_text:
                            text += page_text + "\n"
                    if text.strip():
                        return self.text_processor.clean_text_preserve_newlines(text)
            except Exception as e:
                # If pypdf fails, try PyPDF2 as fallback
                pass
        
        # Fallback to PyPDF2
        try:
            with open(file_path, 'rb') as file:
                # Try to read with strict=False for corrupted PDFs
                try:
                    reader = PyPDF2.PdfReader(file, strict=False)
                except:
                    # If strict=False doesn't work, try normal reader
                    file.seek(0)
                reader = PyPDF2.PdfReader(file)
                
                text = ""
                for i, page in enumerate(reader.pages):
                    try:
                        page_text = page.extract_text()
                        if page_text:
                            text += page_text + "\n"
                    except Exception as page_error:
                        # Skip pages that can't be extracted
                        continue
                
                if not text.strip():
                    raise ValueError(f"No text could be extracted from PDF: {file_path}")
                
                return self.text_processor.clean_text_preserve_newlines(text)
        except Exception as e:
            error_msg = str(e)
            if "EOF marker not found" in error_msg or "EOF" in error_msg:
                raise Exception(
                    f"PDF file appears to be corrupted or incomplete: {file_path}. "
                    f"This may be due to an incomplete upload or corrupted file. "
                    f"Please try re-uploading the file or check if the PDF is valid."
                )
            else:
                raise Exception(f"Error loading PDF {file_path}: {error_msg}")
    
    def load_txt(self, file_path: str) -> str:
        """Load text from TXT file preserving paragraphs"""
        try:
            with open(file_path, 'r', encoding='utf-8') as file:
                text = file.read()
                return self.text_processor.clean_text_preserve_newlines(text)
        except Exception as e:
            raise Exception(f"Error loading TXT {file_path}: {str(e)}")
    
    def load_document(self, file_path: str) -> str:
        """Load document based on file extension"""
        ext = Path(file_path).suffix.lower()
        if ext == '.pdf':
            return self.load_pdf(file_path)
        elif ext == '.txt':
            return self.load_txt(file_path)
        else:
            raise ValueError(f"Unsupported file format: {ext}")

class HierarchyManager:
    """Manage hierarchical metadata definitions"""
    
    def __init__(self, hierarchies_dir: str = "hierarchies"):
        self.hierarchies_dir = Path(hierarchies_dir)
        self.hierarchies = {}
        self.load_hierarchies()
    
    def load_hierarchies(self):
        """Load all hierarchy definitions"""
        for yaml_file in self.hierarchies_dir.glob("*.yaml"):
            with open(yaml_file, 'r', encoding='utf-8') as file:
                hierarchy_name = yaml_file.stem
                self.hierarchies[hierarchy_name] = yaml.safe_load(file)
    
    def get_hierarchy(self, name: str) -> Dict[str, Any]:
        """Get hierarchy definition by name"""
        if name not in self.hierarchies:
            raise ValueError(f"Hierarchy '{name}' not found")
        return self.hierarchies[name]
    
    def list_hierarchies(self) -> List[str]:
        """List available hierarchies"""
        return list(self.hierarchies.keys())

class DocumentChunker:
    """Chunk documents with hierarchical metadata"""
    
    def __init__(self, chunk_size: int = 1000, chunk_overlap: int = 200):
        self.chunk_size = chunk_size
        self.chunk_overlap = chunk_overlap
        self.text_processor = TextProcessor()
        self.hierarchy_manager = HierarchyManager()
        self.ai_detector = OpenAIMetadataDetector(self.hierarchy_manager)
    
    def chunk_document(self, file_path: str, hierarchy: Optional[str], 
                      doc_type: Optional[str], language: Optional[str]) -> List[Chunk]:
        """Chunk document with hierarchical metadata per chunk.
        - Auto-detects hierarchy/doc_type/language when None or 'Auto'.
        - Assigns metadata per chunk to support multi-topic documents.
        """
        loader = DocumentLoader()
        content = loader.load_document(file_path)
        
        # Auto-detect language if needed
        if not language or str(language).lower() == 'auto':
            # Prefer OpenAI if available
            ai_guess = self.ai_detector.detect(content)
            _logger.debug("Language auto-detect: ai_guess=%s", ai_guess.get('language') if isinstance(ai_guess, dict) else None)
            language = ai_guess.get('language') if isinstance(ai_guess, dict) and ai_guess.get('language') in ('en','ja') else (
                'ja' if any('\u3040' <= ch <= '\u30ff' or '\u4e00' <= ch <= '\u9faf' for ch in content) else 'en'
            )

        # Prepare list of hierarchy names and definitions
        hier_names = self.hierarchy_manager.list_hierarchies()

        # If hierarchy is auto, we'll pick best per-chunk later; else load the chosen one
        fixed_hierarchy_def = None
        if hierarchy and hierarchy.lower() != 'auto':
            fixed_hierarchy_def = self.hierarchy_manager.get_hierarchy(hierarchy)

        # Simple structural chunking: split on double newlines first, then fall back to token windows
        raw_blocks = [b.strip() for b in content.split('\n\n') if b.strip()]
        if not raw_blocks:
            raw_blocks = [content]

        # Further split large blocks into overlapping windows
        processed_blocks: List[str] = []
        for block in raw_blocks:
            words = block.split()
            if len(words) <= self.chunk_size:
                processed_blocks.append(block)
            else:
                step = max(1, self.chunk_size - self.chunk_overlap)
                for i in range(0, len(words), step):
                    processed_blocks.append(' '.join(words[i:i + self.chunk_size]))

        # Phase 1: provisional labels for each block
        provisional: List[Dict[str, Any]] = []
        # Sticky explicit labels propagate until overridden by new explicit labels
        sticky_l1: Optional[str] = None
        sticky_l2: Optional[str] = None
        for block in processed_blocks:
            ai_used = False
            ph_hdef = fixed_hierarchy_def
            ph_hname = hierarchy if hierarchy and hierarchy.lower() != 'auto' else None
            if ph_hdef is None:
                ai_guess = self.ai_detector.detect(block)
                guess_name = ai_guess.get('hierarchy_name') if isinstance(ai_guess, dict) else None
                # 0) Explicit label "Hierarchy: <name>"
                import re
                mH = re.search(r"^\s*hierarchy\s*:\s*(.+)$", block, flags=re.IGNORECASE | re.MULTILINE)
                if mH:
                    explicit_h = mH.group(1).strip().lower()
                    for name in hier_names:
                        if name.lower() in explicit_h or explicit_h in name.lower():
                            ph_hdef = self.hierarchy_manager.get_hierarchy(name)
                            ph_hname = name
                            ai_used = ai_used or False
                            
                # 1) If OpenAI guessed a known hierarchy
                if ph_hdef is None and guess_name in hier_names:
                    ph_hdef = self.hierarchy_manager.get_hierarchy(guess_name)
                    ph_hname = guess_name
                    ai_used = True
                # 2) Weighted keyword scoring across all hierarchies (level1/2/3 + doc_types + filename hints)
                if ph_hdef is None:
                    best_score = -1
                    best_name = None
                    best_def = None
                    block_lower = block.lower()
                    filename_lower = os.path.basename(file_path).lower()
                    for name in hier_names:
                        hdef = self.hierarchy_manager.get_hierarchy(name)
                        score = 0
                        # level1
                        for v in hdef['levels']['level1']['values']:
                            if v.lower() in block_lower:
                                score += 2
                        # level2
                        for l2_list in hdef['levels']['level2']['values'].values():
                            for v in l2_list:
                                if v.lower() in block_lower:
                                    score += 2
                        # level3
                        for l3_list in hdef['levels']['level3']['values'].values():
                            for v in l3_list:
                                if v.lower() in block_lower:
                                    score += 1
                        # doc_types
                        for dt in hdef.get('doc_types', []):
                            if dt.lower() in block_lower:
                                score += 1
                        # filename hint
                        if name.lower() in filename_lower:
                            score += 3
                        if score > best_score:
                            best_score = score
                            best_name = name
                            best_def = hdef
                    ph_hdef = best_def if best_def is not None else self.hierarchy_manager.get_hierarchy(hier_names[0])
                    ph_hname = best_name or hier_names[0]

            ph_dtype = doc_type
            if not doc_type or str(doc_type).lower() == 'auto':
                ai_guess = self.ai_detector.detect(block)
                if isinstance(ai_guess, dict) and ai_guess.get('document_type'):
                    ph_dtype = ai_guess['document_type']
                    ai_used = True
                else:
                    dt_candidates = ph_hdef.get('doc_types', ["Policy", "Manual", "FAQ", "Report", "Note", "Guideline"])
                    block_lower = block.lower()
                    best_dt = dt_candidates[0]
                    best_score = -1
                    for dt in dt_candidates:
                        s = 0
                        if dt.lower() in block_lower:
                            s += 1
                        if dt.lower() == 'faq' and ('faq' in block_lower or 'q:' in block_lower):
                            s += 1
                        if dt.lower() == 'report' and ('report' in block_lower or 'summary' in block_lower):
                            s += 1
                        if s > best_score:
                            best_score = s
                            best_dt = dt
                    ph_dtype = best_dt

            content_lower = block.lower()
            # Detect explicit labels in this block
            import re
            exp_l1 = exp_l2 = None
            m1 = re.search(r"^\s*domain\s*:\s*(.+)$", content_lower, flags=re.MULTILINE)
            m2 = re.search(r"^\s*section\s*:\s*(.+)$", content_lower, flags=re.MULTILINE)
            if m1:
                exp_l1 = m1.group(1).strip()
            if m2:
                exp_l2 = m2.group(1).strip()

            # Provisional levels
            ph_l1 = self._classify_level1(content_lower, ph_hdef)
            ph_l2 = self._classify_level2(content_lower, ph_hdef, ph_l1)

            # Override with explicit labels when present
            def _best_match(name: str, candidates: list[str]) -> str:
                name_l = name.lower()
                for c in candidates:
                    cl = c.lower()
                    if cl == name_l or name_l in cl or cl in name_l:
                        return c
                return candidates[0] if candidates else "General"

            if exp_l1:
                ph_l1 = _best_match(exp_l1, ph_hdef['levels']['level1']['values'])
                sticky_l1 = ph_l1
            if exp_l2:
                l2_candidates = ph_hdef['levels']['level2']['values'].get(ph_l1, [])
                ph_l2 = _best_match(exp_l2, l2_candidates)
                sticky_l2 = ph_l2

            # Apply sticky labels when no explicit labels in this block
            if not exp_l1 and sticky_l1:
                ph_l1 = sticky_l1
            if not exp_l2 and sticky_l2 and ph_hdef['levels']['level2']['values'].get(ph_l1):
                ph_l2 = sticky_l2

            provisional.append({
                'text': block,
                'hdef': ph_hdef,
                'hname': ph_hname,
                'dtype': ph_dtype,
                'l1': ph_l1,
                'l2': ph_l2,
                'ai': ai_used
            })

        # Phase 2: merge adjacent blocks with same labels within size limit
        merged_texts: List[str] = []
        merged_meta: List[Dict[str, Any]] = []
        if provisional:
            current_text = provisional[0]['text']
            current_meta = provisional[0]
            for p in provisional[1:]:
                same = (p['hname'] == current_meta['hname'] and p['l1'] == current_meta['l1'] and p['l2'] == current_meta['l2'])
                candidate = current_text + "\n\n" + p['text'] if same else current_text
                if same and self.text_processor.count_tokens(candidate) <= self.text_processor.count_tokens(current_text) + self.chunk_size:
                    current_text = candidate
                    current_meta['ai'] = current_meta['ai'] or p['ai']
                else:
                    merged_texts.append(current_text)
                    merged_meta.append(current_meta)
                    current_text = p['text']
                    current_meta = p
            merged_texts.append(current_text)
            merged_meta.append(current_meta)

        # Phase 3: finalize chunks
        chunks: List[Chunk] = []
        for text_block, meta in zip(merged_texts, merged_meta):
            final_md = self._generate_metadata(
                file_path=file_path,
                hierarchy_def=meta['hdef'],
                doc_type=meta['dtype'],
                language=language,
                content=text_block
            )
            if meta['hname']:
                final_md['hierarchy'] = meta['hname']
            final_md['ai_detected'] = meta['ai']

            chunks.append(Chunk(
                doc_id=generate_id(),
                chunk_id=generate_id(),
                content=text_block,
                metadata=final_md
            ))
        
        return chunks
    
    def _generate_metadata(self, file_path: str, hierarchy_def: Dict[str, Any],
                          doc_type: str, language: str, content: str) -> Dict[str, Any]:
        """Generate hierarchical metadata for chunk"""
        # Simple rule-based classification with explicit label override
        content_lower = content.lower()
        
        # 1) Try to honor explicit labels like "Domain:", "Section:", "Topic:"
        import re
        explicit_l1 = explicit_l2 = explicit_l3 = None
        m1 = re.search(r"^\s*domain\s*:\s*(.+)$", content_lower, flags=re.MULTILINE)
        m2 = re.search(r"^\s*section\s*:\s*(.+)$", content_lower, flags=re.MULTILINE)
        m3 = re.search(r"^\s*topic\s*:\s*(.+)$", content_lower, flags=re.MULTILINE)
        if m1:
            explicit_l1 = m1.group(1).strip()
        if m2:
            explicit_l2 = m2.group(1).strip()
        if m3:
            explicit_l3 = m3.group(1).strip()

        def _best_match(name: str, candidates: list[str]) -> str:
            name_l = name.lower()
            # exact contains
            for c in candidates:
                if c.lower() == name_l or name_l in c.lower() or c.lower() in name_l:
                    return c
            # fallback: first candidate
            return candidates[0] if candidates else "General"

        if explicit_l1:
            level1 = _best_match(explicit_l1, hierarchy_def['levels']['level1']['values'])
        else:
            level1 = self._classify_level1(content_lower, hierarchy_def)

        if explicit_l2:
            level2_candidates = hierarchy_def['levels']['level2']['values'].get(level1, [])
            level2 = _best_match(explicit_l2, level2_candidates)
        else:
            level2 = self._classify_level2(content_lower, hierarchy_def, level1)

        if explicit_l3:
            level3_candidates = hierarchy_def['levels']['level3']['values'].get(level2, [])
            level3 = _best_match(explicit_l3, level3_candidates)
        else:
            level3 = self._classify_level3(content_lower, hierarchy_def, level1, level2)

        # Fallback mapping to 'Other' when nothing matches this hierarchy
        def _any_present(values: list[str]) -> bool:
            return any(v.lower() in content_lower for v in values)

        # If no level1 value appears, set to 'Other'
        if not _any_present(hierarchy_def['levels']['level1']['values']):
            level1 = 'Other'
        # If level2 options for chosen level1 exist but none appear, set to 'Other'
        l2_opts = hierarchy_def['levels']['level2']['values'].get(level1, [])
        if l2_opts and not _any_present(l2_opts):
            level2 = 'Other'
        # If level3 options for chosen level2 exist but none appear, set to 'Other'
        l3_opts = hierarchy_def['levels']['level3']['values'].get(level2, [])
        if l3_opts and not _any_present(l3_opts):
            level3 = 'Other'
        
        return {
            "source_name": os.path.basename(file_path),
            "lang": language,
            "level1": level1,
            "level2": level2,
            "level3": level3,
            "doc_type": doc_type,
            "chunk_size": len(content),
            "token_count": self.text_processor.count_tokens(content)
        }
    
    def _classify_level1(self, content: str, hierarchy_def: Dict[str, Any]) -> str:
        """Classify level1 domain"""
        level1_options = hierarchy_def['levels']['level1']['values']
        
        # Simple keyword matching (enhance with ML model)
        keyword_scores = {}
        for domain in level1_options:
            score = 0
            # Add domain-specific keyword matching logic
            if domain.lower() in content:
                score += 1
            keyword_scores[domain] = score
        
        return max(keyword_scores.items(), key=lambda x: x[1])[0] if keyword_scores else level1_options[0]
    
    def _classify_level2(self, content: str, hierarchy_def: Dict[str, Any], level1: str) -> str:
        """Classify level2 section"""
        level2_options = hierarchy_def['levels']['level2']['values'].get(level1, [])
        if not level2_options:
            return "General"
        
        keyword_scores = {}
        for section in level2_options:
            score = 0
            if section.lower() in content:
                score += 1
            keyword_scores[section] = score
        
        return max(keyword_scores.items(), key=lambda x: x[1])[0] if keyword_scores else level2_options[0]
    
    def _classify_level3(self, content: str, hierarchy_def: Dict[str, Any], 
                        level1: str, level2: str) -> str:
        """Classify level3 topic"""
        level3_options = hierarchy_def['levels']['level3']['values'].get(level2, [])
        if not level3_options:
            return "General"
        
        keyword_scores = {}
        for topic in level3_options:
            score = 0
            if topic.lower() in content:
                score += 1
            keyword_scores[topic] = score
        
        return max(keyword_scores.items(), key=lambda x: x[1])[0] if keyword_scores else level3_options[0]