File size: 15,639 Bytes
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b344aa7
d98f840
 
 
 
 
 
 
 
 
 
 
b344aa7
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f856b7e
 
 
d98f840
f856b7e
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b344aa7
b4d6290
b344aa7
 
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
b344aa7
 
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f856b7e
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b344aa7
 
d98f840
b4d6290
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b4d6290
d98f840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f856b7e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b4d6290
f856b7e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import requests
import logging
from typing import List, Dict, Tuple
import re

logger = logging.getLogger(__name__)

class MedicalReranker:
    """Rerank search results based on medical relevance and source quality"""
    
    def __init__(self):
        self.api_key = os.getenv("NVIDIA_URI")
        self.model = "nvidia/rerank-qa-mistral-4b"
        self.base_url = os.getenv("NVIDIA_RERANK_ENDPOINT", "https://ai.api.nvidia.com/v1/retrieval/nvidia/reranking")
        self.timeout = 30
        
        # Medical domain priority scoring
        self.domain_scores = {
            # High priority medical domains
            'mayoclinic.org': 0.95,
            'webmd.com': 0.90,
            'healthline.com': 0.88,
            'medlineplus.gov': 0.95,
            'nih.gov': 0.98,
            'cdc.gov': 0.98,
            'who.int': 0.97,
            'pubmed.ncbi.nlm.nih.gov': 0.96,
            'uptodate.com': 0.94,
            'merckmanuals.com': 0.92,
            'medscape.com': 0.89,
            'clevelandclinic.org': 0.93,
            'hopkinsmedicine.org': 0.94,
            'harvard.edu': 0.96,
            'stanford.edu': 0.95,
            
            # Medium priority
            'youtube.com': 0.60,  # Lower for general content
            'wikipedia.org': 0.70,
            
            # Low priority (generic health sites)
            'generic_health_site': 0.30
        }
        
        # Irrelevant content patterns - more specific to avoid false positives
        self.irrelevant_patterns = [
            r'quiz|test|assessment|survey',
            r'homepage|main page|index',
            r'login|sign up|register',
            r'contact|about us|privacy',
            r'subscribe|newsletter|rss',
            r'sitemap|search results',
            r'healthtopics\.html',  # Generic topic pages
            r'healthy-sleep/quiz',  # Sleep quiz example
        ]
    
    def rerank_results(self, query: str, results: List[Dict], min_score: float = 0.05) -> List[Dict]:
        """Rerank search results based on medical relevance"""
        if not results:
            return []
        
        # Filter out irrelevant results first
        filtered_results = self._filter_irrelevant_results(results)
        
        if not filtered_results:
            return []
        
        # Score by domain relevance
        domain_scored = self._score_by_domain(filtered_results)
        
        # Use NVIDIA reranker for semantic relevance
        try:
            semantic_scored = self._semantic_rerank(query, domain_scored)
        except Exception as e:
            logger.warning(f"Semantic reranking failed: {e}")
            semantic_scored = domain_scored
        
        # Apply source diversity scoring
        diversity_scored = self._apply_diversity_scoring(semantic_scored)
        
        # Final filtering and sorting
        final_results = [r for r in diversity_scored if r.get('composite_score', 0) >= min_score]
        final_results.sort(key=lambda x: x.get('composite_score', 0), reverse=True)
        
        return final_results
    
    def _filter_irrelevant_results(self, results: List[Dict]) -> List[Dict]:
        """Filter out obviously irrelevant results"""
        filtered = []
        
        for result in results:
            url = result.get('url', '').lower()
            title = result.get('title', '').lower()
            content = result.get('content', '').lower()
            
            # Check for irrelevant patterns
            is_irrelevant = False
            for pattern in self.irrelevant_patterns:
                if re.search(pattern, url) or re.search(pattern, title):
                    is_irrelevant = True
                    break
            
            # Skip if irrelevant
            if is_irrelevant:
                logger.debug(f"Filtered irrelevant result: {url}")
                continue
            
            # Only skip if we have content and it's extremely short
            # Don't filter based on content length if no content is available yet
            if content and len(content) < 20:  # Much more lenient - only filter very short content
                logger.debug(f"Filtered result with very short content: {url}")
                continue
            
            filtered.append(result)
        
        return filtered
    
    def _score_by_domain(self, results: List[Dict]) -> List[Dict]:
        """Score results based on domain credibility"""
        scored_results = []
        
        for result in results:
            url = result.get('url', '')
            domain = self._extract_domain(url)
            
            # Get domain score - be much more lenient with unknown domains
            domain_score = self.domain_scores.get(domain, 0.70)  # Much higher default score
            
            # Boost score for medical-specific content
            title = result.get('title', '').lower()
            content = result.get('content', '').lower()
            
            medical_boost = 0.0
            medical_keywords = [
                'treatment', 'diagnosis', 'symptoms', 'therapy', 'medication',
                'clinical', 'medical', 'health', 'disease', 'condition'
            ]
            
            for keyword in medical_keywords:
                if keyword in title:
                    medical_boost += 0.05
                if keyword in content[:500]:  # Check first 500 chars
                    medical_boost += 0.02
            
            # Calculate composite score
            composite_score = min(domain_score + medical_boost, 1.0)
            
            result['domain_score'] = domain_score
            result['medical_boost'] = medical_boost
            result['composite_score'] = composite_score
            result['domain'] = domain
            
            scored_results.append(result)
        
        return scored_results
    
    def _semantic_rerank(self, query: str, results: List[Dict]) -> List[Dict]:
        """Use NVIDIA reranker for semantic relevance with title prioritization"""
        if not self.api_key:
            return self._fallback_title_rerank(query, results)
        
        # Prepare documents for reranking with title emphasis
        documents = []
        for result in results:
            title = result.get('title', '')
            content = result.get('content', '')[:600]  # Reduced content length
            
            # Prioritize title by repeating it and adding emphasis
            combined_text = f"{title} {title} {content}"  # Title appears twice for emphasis
            documents.append(combined_text)
        
        if not documents:
            return self._fallback_title_rerank(query, results)
        
        try:
            headers = {
                "Authorization": f"Bearer {self.api_key}",
                "Content-Type": "application/json",
            }
            
            payload = {
                "model": self.model,
                "query": query,
                "documents": [{"text": doc} for doc in documents],
            }
            
            response = requests.post(
                self.base_url, 
                headers=headers, 
                json=payload, 
                timeout=self.timeout
            )
            response.raise_for_status()
            data = response.json()
            
            # Process reranking results
            reranked_results = []
            entries = data.get("results") or data.get("data") or []
            
            if entries:
                for i, entry in enumerate(entries):
                    if i < len(results):
                        result = results[i].copy()
                        semantic_score = entry.get("score", 0.5)
                        
                        # Enhanced scoring with title relevance boost
                        domain_score = result.get('domain_score', 0.3)
                        title_relevance = self._calculate_title_relevance(query, result.get('title', ''))
                        
                        # Weighted combination: 30% domain, 50% semantic, 20% title relevance
                        final_score = (domain_score * 0.3) + (semantic_score * 0.5) + (title_relevance * 0.2)
                        
                        result['semantic_score'] = semantic_score
                        result['title_relevance'] = title_relevance
                        result['composite_score'] = final_score
                        reranked_results.append(result)
            else:
                # Fallback to title-based reranking
                reranked_results = self._fallback_title_rerank(query, results)
            
            return reranked_results
            
        except Exception as e:
            logger.warning(f"NVIDIA reranking failed: {e}")
            return self._fallback_title_rerank(query, results)
    
    def _fallback_title_rerank(self, query: str, results: List[Dict]) -> List[Dict]:
        """Fallback reranking based on title relevance when NVIDIA API fails"""
        query_words = set(query.lower().split())
        
        for result in results:
            title = result.get('title', '').lower()
            title_words = set(title.split())
            
            # Calculate title relevance score
            if query_words and title_words:
                overlap = len(query_words.intersection(title_words))
                title_relevance = overlap / len(query_words)
            else:
                title_relevance = 0.0
            
            # Boost for exact phrase matches
            if query.lower() in title:
                title_relevance = min(title_relevance + 0.3, 1.0)
            
            # Update composite score - be much more lenient
            domain_score = result.get('domain_score', 0.7)  # Much higher default
            result['title_relevance'] = title_relevance
            result['composite_score'] = (domain_score * 0.3) + (title_relevance * 0.7)  # Favor title relevance
        
        return results
    
    def _calculate_title_relevance(self, query: str, title: str) -> float:
        """Calculate relevance score based on title and query matching"""
        if not title or not query:
            return 0.0
        
        query_lower = query.lower()
        title_lower = title.lower()
        
        # Exact phrase match gets highest score
        if query_lower in title_lower:
            return 1.0
        
        # Word overlap scoring
        query_words = set(query_lower.split())
        title_words = set(title_lower.split())
        
        if not query_words:
            return 0.0
        
        # Calculate overlap ratio
        overlap = len(query_words.intersection(title_words))
        base_score = overlap / len(query_words)
        
        # Boost for medical terms in title
        medical_terms = ['treatment', 'diagnosis', 'symptoms', 'therapy', 'medication', 'medical', 'health']
        medical_boost = sum(0.15 for term in medical_terms if term in title_lower)
        
        return min(base_score + medical_boost, 1.0)
    
    def _extract_domain(self, url: str) -> str:
        """Extract domain from URL"""
        try:
            from urllib.parse import urlparse
            parsed = urlparse(url)
            domain = parsed.netloc.lower()
            
            # Remove www prefix
            if domain.startswith('www.'):
                domain = domain[4:]
            
            return domain
        except:
            return 'unknown'
    
    def filter_youtube_results(self, results: List[Dict], query: str) -> List[Dict]:
        """Filter and improve YouTube results for medical queries"""
        filtered = []
        
        for result in results:
            url = result.get('url', '')
            title = result.get('title', '')
            
            # Skip generic YouTube search pages
            if 'results?search_query=' in url:
                continue
            
            # Skip non-medical content
            if not self._is_medical_video(title, query):
                continue
            
            # Extract video ID and create proper URL
            video_id = self._extract_video_id(url)
            if video_id:
                result['url'] = f"https://www.youtube.com/watch?v={video_id}"
                result['video_id'] = video_id
                result['source_type'] = 'video'
                filtered.append(result)
        
        return filtered
    
    def _is_medical_video(self, title: str, query: str) -> bool:
        """Check if video title is medically relevant"""
        title_lower = title.lower()
        query_lower = query.lower()
        
        # Medical keywords
        medical_keywords = [
            'medical', 'health', 'doctor', 'treatment', 'diagnosis',
            'symptoms', 'therapy', 'medicine', 'clinical', 'patient'
        ]
        
        # Check if title contains medical keywords or query terms
        has_medical = any(keyword in title_lower for keyword in medical_keywords)
        has_query = any(word in title_lower for word in query_lower.split())
        
        return has_medical or has_query
    
    def _extract_video_id(self, url: str) -> str:
        """Extract YouTube video ID from URL"""
        patterns = [
            r'(?:v=|\/)([0-9A-Za-z_-]{11}).*',
            r'(?:embed\/)([0-9A-Za-z_-]{11})',
            r'(?:watch\?v=)([0-9A-Za-z_-]{11})'
        ]
        
        for pattern in patterns:
            match = re.search(pattern, url)
            if match:
                return match.group(1)
        
        return None
    
    def _apply_diversity_scoring(self, results: List[Dict]) -> List[Dict]:
        """Apply diversity scoring to avoid too many results from same domain"""
        if not results:
            return results
        
        from urllib.parse import urlparse
        from collections import defaultdict
        
        # Track domain counts
        domain_counts = defaultdict(int)
        max_per_domain = 3  # Maximum results per domain
        
        # Apply diversity penalty
        for result in results:
            url = result.get('url', '')
            try:
                domain = urlparse(url).netloc.lower()
                if domain.startswith('www.'):
                    domain = domain[4:]
                
                # Count current domain usage
                domain_counts[domain] += 1
                
                # Apply penalty if domain is over-represented
                if domain_counts[domain] > max_per_domain:
                    # Reduce score for over-represented domains
                    current_score = result.get('composite_score', 0)
                    penalty = 0.15 * (domain_counts[domain] - max_per_domain)
                    result['composite_score'] = max(0, current_score - penalty)
                    result['diversity_penalty'] = penalty
                    logger.debug(f"Applied diversity penalty {penalty} to {domain}")
                else:
                    result['diversity_penalty'] = 0
                    
            except Exception as e:
                logger.debug(f"Error parsing domain for diversity scoring: {e}")
                result['diversity_penalty'] = 0
        
        # Log diversity statistics
        total_domains = len(domain_counts)
        over_represented = sum(1 for count in domain_counts.values() if count > max_per_domain)
        logger.info(f"Diversity scoring: {total_domains} domains, {over_represented} over-represented")
        
        return results