feat: implement 'latest=3 days' logic and Jina content date extraction

src/core/orchestrator/query_orchestrator.py

@@ -67,7 +67,7 @@ class QueryOrchestrator:
     TEMPORAL_KEYWORDS = [
         "today", "now", "latest", "breaking", "just", "current",
         "this morning", "this afternoon", "this evening", "tonight",
-        "yesterday", "recent", "recently", "new", "fresh"
+        "yesterday", "recent", "recently", "new", "fresh", "update"
     ]
     
     # Historical keywords that indicate DB-only search

src/core/use_cases/rag_chat_use_case.py

@@ -162,6 +162,19 @@ Document:
         
         return best_sentence.strip()[:250]
 
+    def _try_extract_date_from_content(self, content: str) -> Optional[str]:
+        """
+        Attempts to find a date string (YYYY-MM-DD) in the first 500 chars of content.
+        Useful for Jina-extracted content where metadata might be missing.
+        """
+        import re
+        # Look for YYYY-MM-DD or Month DD, YYYY
+        iso_match = re.search(r'(\d{4}-\d{2}-\d{2})', content[:500])
+        if iso_match:
+            return iso_match.group(1)
+        
+        return None
+
     def _extract_intents_and_translate(self, query: str) -> Dict[str, Any]:
         """
         Single LLM call: query understanding + multilingual translation.
@@ -272,8 +285,10 @@ JSON:"""
         query_lower = query.lower()
 
         # Temporal fallback - more aggressive keywords
-        if any(w in query_lower for w in ("today", "tonight", "now", "breaking", "latest", "current", "update")):
+        if any(w in query_lower for w in ("today", "tonight", "now", "breaking", "just")):
             result["days_back"] = 1
+        elif any(w in query_lower for w in ("latest", "recently", "current", "update")):
+            result["days_back"] = 3  # "Latest" = 3 days as requested
         elif "yesterday" in query_lower:
             result["days_back"] = 2
         elif any(w in query_lower for w in ("this week", "recently", "past few days", "days ago")):
@@ -331,16 +346,26 @@ JSON:"""
             for hit in results:
                 score_multiplier = 1.0
                 pub_at = hit.metadata.get("published_at")
+                
+                # ENHANCEMENT: If no date in metadata, try to extract from content (Jina fallback)
+                if not pub_at:
+                    pub_at = self._try_extract_date_from_content(hit.content)
+                
                 if pub_at:
                     try:
-                        pub_date = datetime.fromisoformat(pub_at.replace("Z", "+00:00"))
+                        # Handle various date formats (ISO, or simple YYYY-MM-DD)
+                        if "T" in pub_at:
+                            pub_date = datetime.fromisoformat(pub_at.replace("Z", "+00:00"))
+                        else:
+                            pub_date = datetime.strptime(pub_at[:10], "%Y-%m-%d")
+                        
                         days_old = (now.replace(tzinfo=None) - pub_date.replace(tzinfo=None)).days
                         # Aggressive news freshness multiplier
-                        if days_old <= 0: score_multiplier = 1.05   # Boost today
-                        elif days_old == 1: score_multiplier = 1.0  # Normal yesterday
+                        if days_old <= 0: score_multiplier = 1.1   # Even more boost for today
+                        elif days_old <= 3: score_multiplier = 1.0  # Full score for "latest" (3 days)
                         elif days_old < 7: score_multiplier = 0.85  # Slight penalty this week
-                        elif days_old < 30: score_multiplier = 0.6  # Heavy penalty this month
-                        else: score_multiplier = 0.3                 # Archive data
+                        elif days_old < 30: score_multiplier = 0.5  # Heavy penalty this month
+                        else: score_multiplier = 0.2                 # Archive data
                     except:
                         pass