Spaces:
Running
Running
File size: 19,646 Bytes
ddace5c | 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 | """A3 — Reference Checker Agent.
Cross-references a claim against:
1. **RSS feed corpus** — 10 trusted Arabic news outlets (cached for 10 min).
2. **Live web search** — DuckDuckGo + SearXNG (+ Brave/Tavily/Serper if keys),
plus Wikipedia, plus site-targeted searches on a list of trusted Arabic news
domains. This makes the agent able to find a claim *even when only one
source on the web mentions it*.
3. **Article scraping** — top web hits are pulled and parsed with trafilatura
so the similarity ranker has full article text, not just snippets.
The agent merges everything, computes semantic similarity with sentence-
transformers, and classifies into matches / contradictions / web corroboration.
"""
from __future__ import annotations
import asyncio
import logging
import os
import re
import time
from datetime import datetime, timezone
from typing import Optional
from models.response_models import AgentResult, ConfidenceLevel
from services.query_expander import QueryExpander
from services.rss_crawler import Article, RSSCrawler
from services.web_scraper import ScrapedArticle, WebScraper
from services.web_search import SearchResult, WebSearchService
logger = logging.getLogger(__name__)
SOURCE_CREDIBILITY: dict[str, float] = {
"petra.gov.jo": 0.98,
"reuters.com": 0.96,
"apnews.com": 0.96,
"bbci.co.uk": 0.95,
"bbc.co.uk": 0.95,
"bbc.com": 0.95,
"afp.com": 0.94,
"aljazeera.net": 0.92,
"aljazeera.com": 0.92,
"alarabiya.net": 0.88,
"skynewsarabia.com": 0.86,
"france24.com": 0.86,
"dw.com": 0.86,
"alhurra.com": 0.84,
"alghad.com": 0.80,
"alrai.com": 0.78,
"asharq.com": 0.78,
"rt.com": 0.65,
"arabic.rt.com": 0.65,
"jo24.net": 0.65,
"ar.wikipedia.org": 0.78,
"en.wikipedia.org": 0.80,
}
DEFAULT_CREDIBILITY = 0.55
STRONG_MATCH = 0.55 # any source: similarity at/above is a strong match
WEB_FOUND_MATCH = 0.28 # web hits down to here surface as "found on web"
WEAK_MATCH = 0.22 # contradictions / loose-overlap floor
OLD_NEWS_DAYS = 180
ARABIC_STOPWORDS = {
"في", "من", "إلى", "على", "عن", "هو", "هي", "هم", "هذا", "هذه",
"ذلك", "تلك", "أن", "إن", "لا", "ما", "كان", "كانت", "قد", "لقد",
"كل", "بعض", "غير", "أو", "ثم", "أيضاً",
}
class ReferenceCheckerAgent:
"""A3 — multi-source semantic cross-reference."""
def __init__(
self,
timeout_seconds: float = 28.0,
crawler: Optional[RSSCrawler] = None,
web_search: Optional[WebSearchService] = None,
scraper: Optional[WebScraper] = None,
query_expander: Optional[QueryExpander] = None,
) -> None:
self.agent_id = "a3_reference"
self.timeout = timeout_seconds
self.crawler = crawler or RSSCrawler()
self.web_search = web_search or WebSearchService()
self.scraper = scraper or WebScraper()
self.query_expander = query_expander or QueryExpander()
self._encoder = None
self._encoder_failed = False
self.hf_token = os.getenv("HF_API_TOKEN", "").strip()
self._article_cache: list[Article] = []
self._cache_at: Optional[datetime] = None
self._cache_ttl_seconds = 600
self._cache_lock = asyncio.Lock()
async def analyze(self, text: str) -> AgentResult:
start = time.perf_counter()
try:
return await asyncio.wait_for(self._analyze(text, start), timeout=self.timeout)
except asyncio.TimeoutError:
return self._timeout(start)
except Exception as e:
logger.exception("A3 error: %s", e)
return self._fallback(text, start, error=str(e))
async def _analyze(self, text: str, start: float) -> AgentResult:
text = (text or "").strip()
if len(text) < 5:
return self._fallback(text, start, error="text too short")
rss_task = asyncio.create_task(self._get_articles())
queries_task = asyncio.create_task(self.query_expander.expand(text, max_queries=3))
rss_articles, queries = await asyncio.gather(rss_task, queries_task)
if not queries:
queries = [text[:200]]
# Run a single comprehensive web search using the strongest query.
# comprehensive_search itself already fans out (1 main + 2 site +
# Wikipedia) and dedupes, so we don't need additional shadow queries.
primary_query = queries[0]
is_arabic = sum(1 for c in text if "" <= c <= "ۿ") > len(text) * 0.2
lang = "ar" if is_arabic else "en"
web_hits = await self._safe_web_search(primary_query, language=lang)
scraped = await self._scrape_hits(web_hits)
corpus = self._build_corpus(rss_articles, web_hits, scraped)
if not corpus:
return AgentResult(
agent=self.agent_id,
score=0.5,
confidence=ConfidenceLevel.INSUFFICIENT_DATA,
evidence=["no corpus available (RSS+web both empty)"],
raw={"matches": [], "contradictions": [], "web_results": [], "queries": queries},
elapsed_ms=int((time.perf_counter() - start) * 1000),
mode="fallback",
)
sims, sim_mode = await self._similarity(text, corpus)
matches: list[dict] = []
contradictions: list[dict] = []
web_results: list[dict] = []
rss_match_count = 0
web_match_count = 0
for similarity, item in sims:
entry = self._serialize(item, similarity)
if item["kind"] == "web":
if similarity >= WEB_FOUND_MATCH:
web_results.append(entry)
if similarity >= STRONG_MATCH:
matches.append(entry)
web_match_count += 1
elif similarity >= WEAK_MATCH:
contradictions.append(entry)
else: # rss
if similarity >= STRONG_MATCH:
matches.append(entry)
rss_match_count += 1
elif similarity >= WEAK_MATCH:
contradictions.append(entry)
matches.sort(key=lambda m: m["similarity"], reverse=True)
contradictions.sort(key=lambda m: m["similarity"], reverse=True)
web_results.sort(key=lambda m: m["similarity"], reverse=True)
matches = matches[:8]
contradictions = contradictions[:6]
web_results = web_results[:10]
if matches:
credibility = sum(
SOURCE_CREDIBILITY.get(m.get("domain") or "", DEFAULT_CREDIBILITY)
for m in matches
) / len(matches)
score = round(credibility, 4)
elif web_results:
top_sim = web_results[0]["similarity"]
score = round(min(0.45, max(0.25, top_sim * 0.6)), 4)
credibility = score
else:
credibility = 0.0
score = round(min(0.20, max(s for s, _ in sims) if sims else 0.0), 4)
date_ok = self._check_date_ok(matches)
confidence = self._classify_confidence(matches, web_results, sim_mode)
evidence: list[str] = []
for m in matches[:3]:
evidence.append(f"match:{m['source_name']}|sim={m['similarity']:.2f}")
if web_match_count and not matches:
evidence.append(f"web_only_match:{web_match_count}")
if web_results and not matches:
top = web_results[0]
evidence.append(f"web_corroboration:{top['source_name']}|sim={top['similarity']:.2f}")
if contradictions:
evidence.append(f"weak_or_contradiction:{len(contradictions)}")
if not matches and not web_results:
evidence.append("no source mentions this claim")
if not date_ok and matches:
evidence.append("matches older than 180 days")
return AgentResult(
agent=self.agent_id,
score=score,
confidence=confidence,
evidence=evidence[:10],
raw={
"matches": matches,
"contradictions": contradictions,
"web_results": web_results,
"queries": queries,
"engines": self.web_search.status(),
"rss_articles_checked": len(rss_articles),
"web_hits_seen": len(web_hits),
"web_articles_scraped": sum(1 for s in scraped if s.text),
"credibility_score": round(credibility, 4),
"date_ok": date_ok,
"rss_match_count": rss_match_count,
"web_match_count": web_match_count,
},
elapsed_ms=int((time.perf_counter() - start) * 1000),
mode=sim_mode,
)
# -------- corpus + scoring --------
async def _safe_web_search(self, query: str, language: str = "ar") -> list[SearchResult]:
try:
return await self.web_search.comprehensive_search(
query, language=language, per_query_limit=5, total_limit=18,
)
except Exception as e:
logger.warning("comprehensive_search failed: %s", e)
return []
async def _scrape_hits(self, hits: list[SearchResult]) -> list[ScrapedArticle]:
if not hits:
return []
urls = [h.url for h in hits[:5] if h.url]
try:
scraped = await asyncio.wait_for(
self.scraper.fetch_many(urls, concurrency=4),
timeout=8.0,
)
# discard failed extractions (no text)
return [s for s in scraped if s.text and s.word_count >= 25]
except Exception as e:
logger.warning("Scrape phase failed/timed-out: %s", e)
return []
def _build_corpus(
self,
rss_articles: list[Article],
web_hits: list[SearchResult],
scraped: list[ScrapedArticle],
) -> list[dict]:
corpus: list[dict] = []
for a in rss_articles:
text = (a.summary or a.text or "")[:1200]
if not (a.title or text):
continue
corpus.append({
"title": a.title,
"text": (a.title + ". " + text).strip(". "),
"url": a.url,
"source_name": a.source_name,
"domain": a.domain,
"pub_date": a.pub_date.isoformat() if a.pub_date else None,
"kind": "rss",
"engine": "rss",
})
# Map scraped by canonical URL for fast lookup
scraped_by_url: dict[str, ScrapedArticle] = {}
for s in scraped:
scraped_by_url[s.canonical_url or s.url] = s
seen_urls = {a.url for a in rss_articles if a.url}
for hit in web_hits:
url = hit.url
if not url or url in seen_urls:
continue
seen_urls.add(url)
sa = scraped_by_url.get(url) or scraped_by_url.get(self._canonicalize(url))
title = (sa.title if sa else "") or hit.title
domain = (sa.domain if sa else "") or hit.source_domain
pub_date = (sa.pub_date if sa else None) or hit.pub_date
# Build a focused similarity corpus from the most relevant slice:
# search-engine snippet first (engine already deemed it relevant
# to the query), then title and a short prefix of the scraped
# article so we don't drown in boilerplate.
engine_blurb = (hit.snippet or "").strip()
scraped_prefix = (sa.text[:600].strip() if sa and sa.text else "")
sim_text = (title + ". " + engine_blurb + ". " + scraped_prefix).strip(". ").strip()
if not sim_text:
continue
display_summary = (
(sa.summary[:300] if sa and sa.summary else "")
or engine_blurb[:300]
)
corpus.append({
"title": title,
"text": sim_text[:1500],
"summary": display_summary,
"url": url,
"source_name": domain or hit.engine,
"domain": domain,
"pub_date": pub_date,
"kind": "web",
"engine": hit.engine,
})
return corpus
@staticmethod
def _canonicalize(url: str) -> str:
try:
from urllib.parse import urlparse
p = urlparse(url)
host = p.netloc.lower()
if host.startswith("www."):
host = host[4:]
return f"{p.scheme}://{host}{p.path.rstrip('/')}"
except Exception:
return url
async def _similarity(
self, claim: str, corpus: list[dict]
) -> tuple[list[tuple[float, dict]], str]:
if self.hf_token:
via_api = await self._similarity_hf_api(claim, corpus)
if via_api:
return via_api, "real"
encoder = await self._get_encoder()
if encoder is not None:
try:
loop = asyncio.get_running_loop()
texts = [claim] + [c["text"] for c in corpus]
vectors = await loop.run_in_executor(
None,
lambda: encoder.encode(texts, normalize_embeddings=True),
)
claim_vec = vectors[0]
pairs: list[tuple[float, dict]] = []
for vec, item in zip(vectors[1:], corpus):
sim = float((claim_vec * vec).sum())
pairs.append((sim, item))
return pairs, "real"
except Exception as e:
logger.warning("Local encoder failed: %s", e)
return self._jaccard_fallback(claim, corpus), "fallback"
async def _similarity_hf_api(
self, claim: str, corpus: list[dict]
) -> Optional[list[tuple[float, dict]]]:
try:
import httpx
async with httpx.AsyncClient(timeout=10.0) as client:
resp = await client.post(
"https://api-inference.huggingface.co/models/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
headers={"Authorization": f"Bearer {self.hf_token}"},
json={
"inputs": {
"source_sentence": claim[:480],
"sentences": [c["text"][:480] for c in corpus],
}
},
)
if resp.status_code != 200:
return None
scores = resp.json()
if not isinstance(scores, list):
return None
return list(zip([float(s) for s in scores], corpus))
except Exception as e:
logger.debug("HF inference API failed: %s", e)
return None
async def _get_encoder(self):
if self._encoder is not None or self._encoder_failed:
return self._encoder
loop = asyncio.get_running_loop()
try:
self._encoder = await loop.run_in_executor(None, self._load_encoder)
except Exception as e:
logger.warning("Sentence-transformer load failed: %s", e)
self._encoder_failed = True
self._encoder = None
return self._encoder
@staticmethod
def _load_encoder():
from sentence_transformers import SentenceTransformer
return SentenceTransformer("paraphrase-multilingual-MiniLM-L12-v2")
@staticmethod
def _jaccard_fallback(claim: str, corpus: list[dict]) -> list[tuple[float, dict]]:
ctokens = ReferenceCheckerAgent._tokens(claim)
out = []
if not ctokens:
return out
for c in corpus:
t = ReferenceCheckerAgent._tokens(c["text"])
if not t:
continue
jaccard = len(ctokens & t) / max(1, len(ctokens | t))
out.append((jaccard * 0.85, c))
return out
@staticmethod
def _tokens(text: str) -> set[str]:
text = re.sub(r"[^-ۿa-zA-Z0-9 ]", " ", text or "")
return {
t.strip()
for t in text.split()
if t.strip() and len(t.strip()) > 2 and t.strip() not in ARABIC_STOPWORDS
}
@staticmethod
def _serialize(item: dict, similarity: float) -> dict:
return {
"title": item.get("title") or "",
"summary": (item.get("summary") or item.get("text") or "")[:300],
"url": item.get("url") or "",
"source_name": item.get("source_name") or item.get("domain") or "",
"domain": item.get("domain") or "",
"pub_date": item.get("pub_date"),
"kind": item.get("kind") or "",
"engine": item.get("engine") or "",
"similarity": round(float(similarity), 4),
}
async def _get_articles(self) -> list[Article]:
async with self._cache_lock:
now = datetime.now(timezone.utc)
if (
self._article_cache
and self._cache_at
and (now - self._cache_at).total_seconds() < self._cache_ttl_seconds
):
return self._article_cache
try:
articles = await self.crawler.fetch_all(limit_per_source=15)
if articles:
self._article_cache = articles
self._cache_at = now
return articles
except Exception as e:
logger.warning("RSS fetch failed in A3: %s", e)
return self._article_cache
@staticmethod
def _check_date_ok(matches: list[dict]) -> bool:
if not matches:
return False
now = datetime.now(timezone.utc)
for m in matches:
iso = m.get("pub_date")
if not iso:
continue
try:
dt = datetime.fromisoformat(str(iso).replace("Z", "+00:00"))
if (now - dt).days <= OLD_NEWS_DAYS:
return True
except Exception:
continue
return False
@staticmethod
def _classify_confidence(
matches: list[dict], web_results: list[dict], mode: str
) -> ConfidenceLevel:
if not matches and not web_results:
return ConfidenceLevel.LOW
if mode == "real" and len(matches) >= 2:
return ConfidenceLevel.HIGH
if mode == "real" and (matches or len(web_results) >= 2):
return ConfidenceLevel.MEDIUM
return ConfidenceLevel.LOW
def _timeout(self, start: float) -> AgentResult:
return AgentResult(
agent=self.agent_id,
score=0.5,
confidence=ConfidenceLevel.TIMEOUT,
evidence=["agent timeout"],
raw={"matches": [], "contradictions": [], "web_results": []},
elapsed_ms=int((time.perf_counter() - start) * 1000),
mode="fallback",
)
def _fallback(self, text: str, start: float, error: str = "") -> AgentResult:
return AgentResult(
agent=self.agent_id,
score=0.5,
confidence=ConfidenceLevel.INSUFFICIENT_DATA,
evidence=[f"error:{error[:80]}"] if error else [],
raw={"matches": [], "contradictions": [], "web_results": [], "fallback": True},
elapsed_ms=int((time.perf_counter() - start) * 1000),
mode="fallback",
)
|