File size: 32,561 Bytes
4b17916
2a0098d
6142af3
 
46a015f
132c134
6c6c904
 
4d61bdb
6c6c904
6142af3
ffce11c
6142af3
2a0098d
4b17916
2a0098d
6c6c904
4d61bdb
4b17916
 
6c6c904
 
 
 
 
2a0098d
 
6c6c904
2a0098d
 
0e14740
6ac9507
2a0098d
46a015f
4c88f38
31e12c0
4118768
 
 
 
 
4d61bdb
46a015f
6c6c904
 
 
 
 
 
 
 
 
 
 
 
4b17916
6c6c904
 
 
 
 
1e679fd
6142af3
 
4118768
6142af3
ffce11c
 
4118768
 
6c6c904
 
 
 
 
 
 
ffce11c
 
132c134
6c6c904
4118768
 
 
132c134
6c6c904
 
 
 
132c134
 
6c6c904
 
b7afcad
 
4d61bdb
b7afcad
 
 
6c6c904
4d61bdb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4118768
 
 
 
 
 
4d61bdb
 
 
 
 
 
 
 
 
 
 
 
 
6c6c904
 
 
b7afcad
 
6c6c904
 
b7afcad
6c6c904
 
 
 
 
 
 
 
 
b7afcad
6c6c904
b7afcad
6c6c904
 
 
 
b7afcad
6c6c904
 
 
4d61bdb
6c6c904
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d61bdb
 
 
 
 
 
 
 
 
 
 
 
4118768
4d61bdb
 
 
 
 
 
 
 
 
 
6c6c904
4118768
 
 
 
6c6c904
 
4d61bdb
4118768
 
6c6c904
4d61bdb
 
 
6c6c904
 
4d61bdb
6c6c904
 
 
 
 
0eacd1e
6c6c904
0eacd1e
6c6c904
 
4d61bdb
 
 
 
 
6c6c904
 
4d61bdb
 
6c6c904
 
4906187
4d61bdb
6c6c904
 
 
4d61bdb
 
6c6c904
 
 
 
 
4d61bdb
6c6c904
4d61bdb
6c6c904
 
 
 
4d61bdb
6c6c904
 
4906187
 
6c6c904
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d61bdb
 
 
 
6c6c904
 
b7afcad
6c6c904
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d61bdb
b7afcad
4d61bdb
 
 
 
 
 
 
 
 
 
 
 
 
6c6c904
 
 
 
 
 
 
4d61bdb
6c6c904
4906187
4d61bdb
6c6c904
 
 
 
 
 
 
 
 
4d61bdb
 
6c6c904
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4118768
6c6c904
 
4d61bdb
6c6c904
 
b7afcad
4d61bdb
 
 
 
 
6c6c904
 
 
 
 
4d61bdb
 
 
6c6c904
2a0098d
4d61bdb
 
 
4118768
4d61bdb
 
 
 
4118768
 
4d61bdb
 
 
 
 
 
 
 
4118768
 
 
4d61bdb
 
 
 
4118768
4d61bdb
 
 
4118768
 
 
 
4d61bdb
 
 
 
4118768
 
4d61bdb
4118768
 
 
 
 
 
 
 
 
 
 
 
 
 
4d61bdb
4118768
 
 
4d61bdb
 
4118768
4d61bdb
 
 
 
 
4118768
 
4d61bdb
4118768
4d61bdb
 
 
 
4118768
 
 
 
4d61bdb
 
4118768
4d61bdb
 
 
 
 
 
 
 
 
 
 
 
 
4118768
 
 
 
 
4d61bdb
 
 
 
 
 
 
 
 
4118768
4d61bdb
 
4118768
4d61bdb
 
 
 
 
4118768
4d61bdb
 
 
 
4118768
4d61bdb
 
 
 
 
 
 
4118768
 
4d61bdb
 
 
 
 
 
 
4118768
 
4d61bdb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4118768
4d61bdb
4118768
 
 
 
 
 
 
 
 
4d61bdb
4118768
4d61bdb
 
6c6c904
 
 
 
 
 
 
 
6142af3
6c6c904
 
 
4118768
6c6c904
5b2a6b6
6c6c904
 
 
 
4118768
 
 
 
 
 
 
a38a28a
4d61bdb
6c6c904
 
4118768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c6c904
a38a28a
4d61bdb
4118768
 
 
4d61bdb
 
4118768
4d61bdb
6c6c904
4118768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d61bdb
6c6c904
 
 
 
 
 
4d61bdb
4118768
 
 
 
6c6c904
 
4118768
6c6c904
 
4d61bdb
6c6c904
 
4d61bdb
6c6c904
 
 
4118768
 
 
 
 
 
 
 
 
 
 
 
 
 
4d61bdb
6c6c904
 
 
 
 
 
 
 
 
 
 
4d61bdb
6c6c904
 
4d61bdb
6c6c904
 
4118768
6c6c904
 
 
 
 
b7afcad
6c6c904
4118768
 
 
 
 
 
 
 
 
 
 
 
 
6c6c904
 
a38a28a
6c6c904
 
 
 
4118768
0eacd1e
4118768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46a015f
 
6c6c904
4118768
6c6c904
 
4118768
 
 
 
 
 
 
 
46a015f
1e679fd
6c6c904
 
 
 
 
6142af3
4118768
 
 
 
 
 
4d61bdb
 
6c6c904
 
4118768
6c6c904
6142af3
4d61bdb
 
4118768
4d61bdb
 
4118768
4d61bdb
4118768
 
 
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
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
import os
import asyncio
import json
import logging
import random
import re
import time
from typing import AsyncGenerator, Optional, Tuple, List, Dict
from urllib.parse import quote_plus, urlparse
from fastapi import FastAPI, HTTPException
from fastapi.responses import StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from dotenv import load_dotenv
import aiohttp
from bs4 import BeautifulSoup
from fake_useragent import UserAgent
from collections import defaultdict

# --- Configuration ---
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
load_dotenv()

LLM_API_KEY = os.getenv("LLM_API_KEY")
if not LLM_API_KEY:
    raise RuntimeError("LLM_API_KEY must be set in a .env file.")
else:
    logging.info("LLM API Key loaded successfully.")

# --- Constants & Headers ---
LLM_API_URL = "https://api.typegpt.net/v1/chat/completions"
LLM_MODEL = "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
MAX_SOURCES_TO_PROCESS = 10
MAX_CONCURRENT_REQUESTS = 5
SEARCH_TIMEOUT = 120  # Default search time in seconds
TOTAL_TIMEOUT = 180   # Total time limit in seconds
REQUEST_DELAY = 1.0   # Delay between requests in seconds
USER_AGENT_ROTATION = True

# Initialize fake user agent generator
try:
    ua = UserAgent()
except:
    class SimpleUA:
        def random(self):
            return random.choice([
                "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/128.0.0.0 Safari/537.36",
                "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/128.0.0.0 Safari/537.36",
                "Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:129.0) Gecko/20100101 Firefox/129.0"
            ])
    ua = SimpleUA()

LLM_HEADERS = {
    "Authorization": f"Bearer {LLM_API_KEY}",
    "Content-Type": "application/json",
    "Accept": "application/json"
}

class DeepResearchRequest(BaseModel):
    query: str
    search_time: int = SEARCH_TIMEOUT  # Default search time

app = FastAPI(
    title="AI Deep Research API",
    description="Provides comprehensive research reports from real web searches.",
    version="3.1.0"
)
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"]
)

def extract_json_from_llm_response(text: str) -> Optional[list]:
    """Extract JSON array from LLM response text."""
    match = re.search(r'$$
.*
$$', text, re.DOTALL)
    if match:
        try:
            return json.loads(match.group(0))
        except json.JSONDecodeError:
            return None
    return None

async def get_real_user_agent() -> str:
    """Get a realistic user agent string."""
    try:
        if isinstance(ua, UserAgent):
            return ua.random()
        return ua.random()  # For our fallback class
    except:
        return "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/128.0.0.0 Safari/537.36"

def clean_url(url: str) -> str:
    """Clean up and normalize URLs."""
    if not url:
        return ""

    # Handle DuckDuckGo redirect URLs
    if url.startswith('//duckduckgo.com/l/'):
        url = f"https:{url}"  # Make it a proper URL
        try:
            parsed = urlparse(url)
            query_params = parsed.query
            if 'uddg=' in query_params:
                match = re.search(r'uddg=([^&]+)', query_params)
                if match:
                    encoded_url = match.group(1)
                    try:
                        # URL decode the parameter
                        decoded_url = quote_plus(encoded_url)
                        # Sometimes it's double-encoded
                        if '%25' in decoded_url:
                            decoded_url = quote_plus(decoded_url)
                        return decoded_url
                    except:
                        pass
        except:
            pass

    # Ensure URL has proper scheme
    if url.startswith('//'):
        url = 'https:' + url
    elif not url.startswith(('http://', 'https://')):
        url = 'https://' + url

    return url

async def check_robots_txt(url: str) -> bool:
    """Check if scraping is allowed by robots.txt."""
    try:
        domain_match = re.search(r'https?://([^/]+)', url)
        if not domain_match:
            return False

        domain = domain_match.group(1)
        robots_url = f"https://{domain}/robots.txt"

        async with aiohttp.ClientSession() as session:
            headers = {'User-Agent': await get_real_user_agent()}
            async with session.get(robots_url, headers=headers, timeout=5) as response:
                if response.status == 200:
                    robots = await response.text()
                    if "Disallow: /" in robots:
                        return False
                    # Check for specific path disallows
                    path = re.sub(r'https?://[^/]+', '', url)
                    if any(f"Disallow: {p}" in robots for p in [path, path.rstrip('/') + '/']):
                        return False
        return True
    except Exception as e:
        logging.warning(f"Could not check robots.txt for {url}: {e}")
        return False

async def fetch_search_results(query: str, max_results: int = 5) -> List[dict]:
    """
    Perform a real search using DuckDuckGo's HTML interface with improved URL handling.
    """
    try:
        search_url = f"https://html.duckduckgo.com/html/?q={quote_plus(query)}"
        headers = {
            "User-Agent": await get_real_user_agent(),
            "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,*/*;q=0.8",
            "Accept-Language": "en-US,en;q=0.5",
            "Referer": "https://duckduckgo.com/",
            "DNT": "1"
        }

        async with aiohttp.ClientSession() as session:
            async with session.get(search_url, headers=headers, timeout=10) as response:
                if response.status != 200:
                    logging.warning(f"Search failed with status {response.status}")
                    return []

                html = await response.text()
                soup = BeautifulSoup(html, 'html.parser')

                results = []
                # Try multiple selectors as DuckDuckGo may change their HTML structure
                for selector in ['.result__body', '.result__a', '.result']:
                    if len(results) >= max_results:
                        break

                    for result in soup.select(selector)[:max_results]:
                        try:
                            title_elem = result.select_one('.result__title .result__a') or result.select_one('.result__a')
                            if not title_elem:
                                continue

                            link = title_elem['href']
                            snippet_elem = result.select_one('.result__snippet') or result.select_one('.result__body')

                            # Clean the URL
                            clean_link = clean_url(link)

                            # Skip if we couldn't get a clean URL
                            if not clean_link or clean_link.startswith('javascript:'):
                                continue

                            # Get snippet if available
                            snippet = snippet_elem.get_text(strip=True) if snippet_elem else ""

                            # Skip if we already have this URL
                            if any(r['link'] == clean_link for r in results):
                                continue

                            results.append({
                                'title': title_elem.get_text(strip=True),
                                'link': clean_link,
                                'snippet': snippet,
                                'source': 'duckduckgo'
                            })
                        except Exception as e:
                            logging.warning(f"Error parsing search result: {e}")
                            continue

                logging.info(f"Found {len(results)} real search results for '{query}'")
                return results[:max_results]
    except Exception as e:
        logging.error(f"Real search failed: {e}")
        return []

async def process_web_source(session: aiohttp.ClientSession, source: dict, timeout: int = 15) -> Tuple[str, dict]:
    """
    Process a real web source with improved content extraction and error handling.
    """
    headers = {'User-Agent': await get_real_user_agent()}
    source_info = source.copy()
    source_info['link'] = clean_url(source['link'])  # Ensure URL is clean

    # Skip if URL is invalid
    if not source_info['link'] or not source_info['link'].startswith(('http://', 'https://')):
        return source.get('snippet', ''), source_info

    # Check robots.txt first
    if not await check_robots_txt(source_info['link']):
        logging.info(f"Scraping disallowed by robots.txt for {source_info['link']}")
        return source.get('snippet', ''), source_info

    try:
        logging.info(f"Processing source: {source_info['link']}")
        start_time = time.time()

        # Skip non-HTML content
        if any(source_info['link'].lower().endswith(ext) for ext in ['.pdf', '.doc', '.docx', '.ppt', '.pptx', '.xls', '.xlsx']):
            logging.info(f"Skipping non-HTML content at {source_info['link']}")
            return source.get('snippet', ''), source_info

        # Add delay between requests to be polite
        await asyncio.sleep(REQUEST_DELAY)

        async with session.get(source_info['link'], headers=headers, timeout=timeout, ssl=False) as response:
            if response.status != 200:
                logging.warning(f"HTTP {response.status} for {source_info['link']}")
                return source.get('snippet', ''), source_info

            content_type = response.headers.get('Content-Type', '').lower()
            if 'text/html' not in content_type:
                logging.info(f"Non-HTML content at {source_info['link']} (type: {content_type})")
                return source.get('snippet', ''), source_info

            html = await response.text()
            soup = BeautifulSoup(html, "html.parser")

            # Remove unwanted elements
            for tag in soup(['script', 'style', 'nav', 'footer', 'header', 'aside', 'iframe', 'noscript', 'form']):
                tag.decompose()

            # Try to find main content by common patterns
            selectors_to_try = [
                'main',
                'article',
                '[role="main"]',
                '.main-content',
                '.content',
                '.article-body',
                '.post-content',
                '.entry-content',
                '#content',
                '#main',
                '.main',
                '.article'
            ]

            main_content = None
            for selector in selectors_to_try:
                main_content = soup.select_one(selector)
                if main_content:
                    break

            if not main_content:
                # If no main content found, try to find the largest text block
                all_elements = soup.find_all()
                candidates = [el for el in all_elements if el.name not in ['script', 'style', 'nav', 'footer', 'header']]
                if candidates:
                    candidates.sort(key=lambda x: len(x.get_text()), reverse=True)
                    main_content = candidates[0] if candidates else soup

            if not main_content:
                main_content = soup.find('body') or soup

            # Clean up the content
            content = " ".join(main_content.stripped_strings)
            content = re.sub(r'\s+', ' ', content).strip()

            # If content is too short, try alternative extraction methods
            if len(content.split()) < 50 and len(html) > 10000:
                # Try extracting all paragraphs
                paras = soup.find_all('p')
                content = " ".join([p.get_text() for p in paras if p.get_text().strip()])
                content = re.sub(r'\s+', ' ', content).strip()

                # If still too short, try getting all text nodes
                if len(content.split()) < 50:
                    content = " ".join(soup.stripped_strings)
                    content = re.sub(r'\s+', ' ', content).strip()

            # If content is still too short, try to extract from specific tags
            if len(content.split()) < 30:
                # Try to get content from divs with certain classes
                for tag in ['div', 'section', 'article']:
                    for element in soup.find_all(tag):
                        if len(element.get_text().split()) > 200:  # If this element has substantial content
                            content = " ".join(element.stripped_strings)
                            content = re.sub(r'\s+', ' ', content).strip()
                            if len(content.split()) >= 30:  # If we got enough content
                                break
                    if len(content.split()) >= 30:
                        break

            if len(content.split()) < 30:
                logging.warning(f"Very little content extracted from {source_info['link']}")
                return source.get('snippet', ''), source_info

            source_info['word_count'] = len(content.split())
            source_info['processing_time'] = time.time() - start_time
            return content, source_info

    except asyncio.TimeoutError:
        logging.warning(f"Timeout while processing {source_info['link']}")
        return source.get('snippet', ''), source_info
    except Exception as e:
        logging.warning(f"Error processing {source_info['link']}: {str(e)[:200]}")
        return source.get('snippet', ''), source_info

async def generate_research_plan(query: str, session: aiohttp.ClientSession) -> List[str]:
    """Generate a comprehensive research plan with sub-questions."""
    try:
        plan_prompt = {
            "model": LLM_MODEL,
            "messages": [{
                "role": "user",
                "content": f"""Generate 4-6 comprehensive sub-questions for in-depth research on '{query}'.
                Focus on key aspects that would provide a complete understanding of the topic.
                Your response MUST be ONLY the raw JSON array with no additional text.
                Example: ["What is the historical background of X?", "What are the current trends in X?"]"""
            }],
            "temperature": 0.7,
            "max_tokens": 300
        }

        async with session.post(LLM_API_URL, headers=LLM_HEADERS, json=plan_prompt, timeout=30) as response:
            response.raise_for_status()
            result = await response.json()

            if isinstance(result, list):
                return result
            elif isinstance(result, dict) and 'choices' in result:
                content = result['choices'][0]['message']['content']
                sub_questions = extract_json_from_llm_response(content)
                if sub_questions and isinstance(sub_questions, list):
                    cleaned = []
                    for q in sub_questions:
                        if isinstance(q, str) and q.strip():
                            cleaned_q = re.sub(r'^[^a-zA-Z0-9]*|[^a-zA-Z0-9]*\$', '', q)
                            if cleaned_q:
                                cleaned.append(cleaned_q)
                    return cleaned[:6]  # Limit to 6 questions max

        # Fallback if we couldn't get good questions from LLM
        return [
            f"What is {query} and its key features?",
            f"How does {query} compare to alternatives?",
            f"What are the current developments in {query}?",
            f"What are the main challenges with {query}?",
            f"What does the future hold for {query}?"
        ]
    except Exception as e:
        logging.error(f"Failed to generate research plan: {e}")
        return [
            f"What is {query}?",
            f"What are the key aspects of {query}?",
            f"What are current trends in {query}?",
            f"What are the challenges with {query}?"
        ]

async def continuous_search(query: str, search_time: int = 120) -> List[dict]:
    """
    Perform continuous searching for better results within time constraints.
    Provides detailed feedback about the search process.
    """
    start_time = time.time()
    all_results = []
    seen_urls = set()
    seen_domains = defaultdict(int)
    search_iterations = 0

    # Generate multiple variations of the query
    query_variations = [
        query,
        f"{query} comparison",
        f"{query} analysis",
        f"{query} review",
        f"{query} features",
        f"{query} vs alternatives",
        f"latest {query} news",
        f"{query} pros and cons"
    ]

    async with aiohttp.ClientSession() as session:
        while time.time() - start_time < search_time:
            search_iterations += 1
            # Shuffle the query variations to get diverse results
            random.shuffle(query_variations)

            # Use only a subset of queries each iteration
            queries_for_this_iteration = query_variations[:min(3, len(query_variations))]

            for q in queries_for_this_iteration:
                if time.time() - start_time >= search_time:
                    break

                try:
                    # Notify about current search
                    logging.info(f"Searching for: '{q}'")
                    results = await fetch_search_results(q, max_results=5)

                    if results:
                        for result in results:
                            clean_link = clean_url(result['link'])
                            domain = urlparse(clean_link).netloc if clean_link else ""

                            # Skip if we've already seen this URL
                            if clean_link in seen_urls:
                                continue

                            # Skip if we have too many results from this domain
                            if domain and seen_domains[domain] >= 2:  # Max 2 results per domain
                                continue

                            seen_urls.add(clean_link)
                            if domain:
                                seen_domains[domain] += 1

                            result['link'] = clean_link
                            all_results.append(result)
                            logging.info(f"Found new result: {result['title']} ({domain})")

                    # Small delay between searches
                    await asyncio.sleep(1.0)

                    # If we have enough unique results, we can stop early
                    if len(all_results) >= MAX_SOURCES_TO_PROCESS * 2:  # Get more than we need for selection
                        logging.info(f"Found enough unique results ({len(all_results)})")
                        break

                except Exception as e:
                    logging.error(f"Error during continuous search: {e}")
                    await asyncio.sleep(2.0)  # Wait a bit before trying again

            # Break if we've done several iterations
            if search_iterations >= 4:  # Limit to 4 search iterations
                break

    # Filter and sort results by relevance
    if all_results:
        # Simple relevance scoring
        def score_result(result):
            query_terms = set(query.lower().split())
            title = result['title'].lower()
            snippet = result['snippet'].lower()

            matches = 0
            for term in query_terms:
                if term in title or term in snippet:
                    matches += 1

            # Also consider length of snippet as a proxy for content richness
            snippet_length = len(result['snippet'].split())

            # Prefer results from diverse domains
            domain = urlparse(result['link']).netloc if result['link'] else ""
            domain_score = 10 if seen_domains[domain] <= 1 else 5  # Bonus for unique domains

            return matches * 10 + snippet_length + domain_score

        # Sort by score, descending
        all_results.sort(key=lambda x: score_result(x), reverse=True)

    return all_results[:MAX_SOURCES_TO_PROCESS * 2]  # Return more than we need for selection

async def filter_and_select_sources(results: List[dict]) -> List[dict]:
    """
    Filter and select the best sources from search results.
    Returns a tuple of (selected_sources, rejected_sources_with_reasons)
    """
    if not results:
        return [], []

    # Group by domain to ensure diversity
    domain_counts = defaultdict(int)
    domain_results = defaultdict(list)
    for result in results:
        domain = urlparse(result['link']).netloc if result['link'] else ""
        domain_counts[domain] += 1
        domain_results[domain].append(result)

    selected = []
    rejected = []

    # First pass: take the top result from each domain
    for domain, domain_res in domain_results.items():
        if len(selected) >= MAX_SOURCES_TO_PROCESS:
            break
        # Take the best result from this domain (sorted by position in original results)
        if domain_res:
            # Sort domain results by snippet length (proxy for content richness)
            domain_res.sort(key=lambda x: len(x['snippet'].split()), reverse=True)
            selected.append(domain_res[0])

    # Second pass: if we need more, take additional results from domains with good content
    if len(selected) < MAX_SOURCES_TO_PROCESS:
        # Calculate average snippet length as a proxy for content quality
        domain_quality = {}
        for domain, domain_res in domain_results.items():
            if not domain_res:
                continue
            avg_length = sum(len(r['snippet'].split()) for r in domain_res) / len(domain_res)
            domain_quality[domain] = avg_length

        # Sort domains by quality
        sorted_domains = sorted(domain_quality.items(), key=lambda x: x[1], reverse=True)

        # Add more results from high-quality domains
        for domain, _ in sorted_domains:
            if len(selected) >= MAX_SOURCES_TO_PROCESS:
                break
            for res in domain_results[domain]:
                if res not in selected:
                    selected.append(res)
                    if len(selected) >= MAX_SOURCES_TO_PROCESS:
                        break

    # Third pass: if still need more, add remaining high-snippet-length results
    if len(selected) < MAX_SOURCES_TO_PROCESS:
        # Sort all results by snippet length
        remaining_results = [res for res in results if res not in selected]
        remaining_results.sort(key=lambda x: len(x['snippet'].split()), reverse=True)

        while len(selected) < MAX_SOURCES_TO_PROCESS and remaining_results:
            selected.append(remaining_results.pop(0))

    # The remaining results are our rejected ones (for now we won't track reasons)
    rejected = [res for res in results if res not in selected]

    return selected, rejected

async def run_deep_research_stream(query: str, search_time: int = 120) -> AsyncGenerator[str, None]:
    def format_sse(data: dict) -> str:
        return f"data: {json.dumps(data)}\n\n"

    start_time = time.time()
    processed_sources = 0
    successful_sources = 0
    total_tokens = 0

    try:
        # Initialize the SSE stream with start message
        yield format_sse({
            "event": "status",
            "data": f"Starting deep research on '{query}'. Searching for comprehensive sources..."
        })

        async with aiohttp.ClientSession() as session:
            # Step 1: Generate research plan
            yield format_sse({"event": "status", "data": "Generating comprehensive research plan..."})
            sub_questions = await generate_research_plan(query, session)
            yield format_sse({
                "event": "plan",
                "data": {
                    "sub_questions": sub_questions,
                    "message": f"Research will focus on these {len(sub_questions)} key aspects"
                }
            })

            # Step 2: Continuous search for better results
            yield format_sse({
                "event": "status",
                "data": "Performing intelligent search for high-quality sources..."
            })

            # Show search variations we'll use
            query_variations = [
                query,
                f"{query} comparison",
                f"{query} analysis",
                f"{query} review",
                f"{query} features",
                f"{query} vs alternatives"
            ]
            yield format_sse({
                "event": "status",
                "data": f"Using {len(query_variations)} different search variations to find diverse sources"
            })

            search_results = await continuous_search(query, search_time)

            # Report on search results
            unique_domains = len({urlparse(r['link']).netloc for r in search_results if r['link']})
            yield format_sse({
                "event": "status",
                "data": f"Found {len(search_results)} potential sources from {unique_domains} unique domains"
            })

            # Display some of the top sources found
            if search_results:
                top_sources = search_results[:5]  # Show top 5
                sources_list = []
                for i, source in enumerate(top_sources, 1):
                    domain = urlparse(source['link']).netloc if source['link'] else "Unknown"
                    sources_list.append(f"{i}. {source['title']} ({domain})")

                yield format_sse({
                    "event": "sources_found",
                    "data": {
                        "top_sources": sources_list,
                        "total_sources": len(search_results)
                    }
                })

            if not search_results:
                yield format_sse({
                    "event": "error",
                    "data": "No search results found. Check your query and try again."
                })
                return

            # Select the best sources
            selected_sources, rejected_sources = await filter_and_select_sources(search_results)

            # Report on selected sources
            unique_selected_domains = len({urlparse(r['link']).netloc for r in selected_sources if r['link']})
            yield format_sse({
                "event": "status",
                "data": f"Selected {len(selected_sources)} high-quality sources from {unique_selected_domains} unique domains for in-depth analysis"
            })

            if not selected_sources:
                yield format_sse({
                    "event": "error",
                    "data": "No valid sources found after filtering."
                })
                return

            # Show selected sources
            selected_sources_list = []
            for i, source in enumerate(selected_sources, 1):
                domain = urlparse(source['link']).netloc if source['link'] else "Unknown"
                selected_sources_list.append(f"{i}. {source['title']} ({domain})")

            yield format_sse({
                "event": "sources_selected",
                "data": {
                    "selected_sources": selected_sources_list,
                    "message": "Proceeding with in-depth analysis of these sources"
                }
            })

            # Step 3: Process selected sources with concurrency control
            semaphore = asyncio.Semaphore(MAX_CONCURRENT_REQUESTS)
            consolidated_context = ""
            all_sources_used = []
            processing_errors = 0

            async def process_with_semaphore(source):
                async with semaphore:
                    return await process_web_source(session, source, timeout=20)

            # Process sources with progress updates
            processing_tasks = []
            for i, source in enumerate(selected_sources):
                # Check if we're running out of time
                elapsed = time.time() - start_time
                if elapsed > TOTAL_TIMEOUT * 0.8:  # Leave 20% of time for synthesis
                    yield format_sse({
                        "event": "status",
                        "data": f"Approaching time limit, stopping source processing after {i}/{len(selected_sources)} sources"
                    })
                    break

                # Add delay between processing each source to be polite
                if i > 0:
                    await asyncio.sleep(REQUEST_DELAY * 0.5)

                # Notify about processing this source
                domain = urlparse(source['link']).netloc if source['link'] else "Unknown"
                yield format_sse({
                    "event": "processing_source",
                    "data": {
                        "index": i + 1,
                        "total": len(selected_sources),
                        "title": source['title'],
                        "domain": domain,
                        "url": source['link']
                    }
                })

                task = asyncio.create_task(process_with_semaphore(source))
                processing_tasks.append(task)

            # Process completed tasks as they finish
            for future in asyncio.as_completed(processing_tasks):
                processed_sources += 1
                content, source_info = await future

                if content and content.strip():
                    # Report successful processing
                    domain = urlparse(source_info['link']).netloc if source_info['link'] else "Unknown"
                    word_count = len(content.split())

                    yield format_sse({
                        "event": "source_processed",
                        "data": {
                            "title": source_info['title'],
                            "domain": domain,
                            "word_count": word_count,
                            "status": "success"
                        }
                    })

                    # Add to our consolidated context
                    consolidated_context += f"Source: {source_info['link']}\nContent: {content}\n\n---\n\n"
                    all_sources_used.append(source_info)
                    successful_sources += 1
                    total_tokens += word_count  # Add to token count
                else:
                    processing_errors += 1
                    yield format_sse({
                        "event": "source_processed",
                        "data": {
                            "title": source_info['title'],
                            "status": "failed",
                            "reason": "Could not extract sufficient content"
                        }
                    })

            if not consolidated_context.strip():
                yield format_sse({
                    "event": "error",
                    "data": f"Failed to extract content from any sources. {processing_errors} errors occurred."
                })
                return

            # Report on processing results
            yield format_sse({
                "event": "status",
                "data": f"Successfully processed {successful_sources} of {processed_sources} sources, extracting approximately {total_tokens} words of content"
            })

            # Step 4: Synthesize comprehensive report
            time_remaining = max(0, TOTAL_TIMEOUT - (time.time() - start_time))
            yield format_sse({
                "event": "status",
                "data": f"Generating comprehensive analysis report from {successful_sources} sources..."
            })

            max_output_tokens = min(2000, int(time_remaining * 6))  # More aggressive token count

            report_prompt = f"""Compose a comprehensive analysis report on "{query}".

            Structure the report with these sections:
            1. Executive Summary
            2. Key Features and Capabilities
            3. Comparative Analysis
            4. Strengths and Weaknesses
            5. Current Trends and