fix(nordagpt): structural anti-hallucination — validate ALL company links against DB

Prompt rules don't work — AI ignores them and invents company names.
Added _validate_company_references() post-processor that:
- Loads all valid company slugs from DB
- Scans every /firma/ link in AI response
- REMOVES links to companies that don't exist
- Cleans up empty list items left by removals
- Applied to BOTH send_message() and send_message_stream()

This is the ONLY reliable way to prevent hallucinated companies.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

nordabiz_chat.py

@@ -134,6 +134,61 @@ class NordaBizChatEngine:
             self.model_name = self.gemini_service.model_name
             self.model = None
 
+    @staticmethod
+    def _get_valid_company_slugs() -> set:
+        """Load all valid company slugs from DB. Cached per-request."""
+        db = SessionLocal()
+        try:
+            from database import Company
+            rows = db.query(Company.slug, Company.name).filter(Company.status == 'active').all()
+            return {r.slug: r.name for r in rows if r.slug}
+        finally:
+            db.close()
+
+    @staticmethod
+    def _validate_company_references(text: str) -> str:
+        """
+        Post-process AI response: remove links to companies that don't exist in DB.
+        This is the ONLY reliable way to prevent hallucinated company names.
+        """
+        import re
+
+        valid_companies = NordaBizChatEngine._get_valid_company_slugs()
+        valid_slugs = set(valid_companies.keys())
+        valid_names_lower = {name.lower(): name for name in valid_companies.values()}
+
+        # 1. Validate markdown links to /firma/slug — remove if slug doesn't exist
+        def replace_link(match):
+            link_text = match.group(1)
+            slug = match.group(2)
+            if slug in valid_slugs:
+                return match.group(0)  # Keep valid link
+            else:
+                logger.warning(f"NordaGPT hallucination blocked: removed link to non-existent company slug '{slug}' (text: '{link_text}')")
+                return ''  # Remove entire link
+
+        text = re.sub(r'\[([^\]]+)\]\(/firma/([a-z0-9-]+)\)', replace_link, text)
+
+        # 2. Validate pill-style links that the frontend renders
+        def replace_pill_link(match):
+            full_match = match.group(0)
+            slug = match.group(1)
+            if slug in valid_slugs:
+                return full_match
+            else:
+                logger.warning(f"NordaGPT hallucination blocked: removed pill link to '{slug}'")
+                return ''
+
+        text = re.sub(r'<a[^>]*href=["\']/firma/([a-z0-9-]+)["\'][^>]*>.*?</a>', replace_pill_link, text)
+
+        # 3. Clean up empty list items and double spaces left by removals
+        text = re.sub(r'\n\s*\*\s*\n', '\n', text)  # empty bullet points
+        text = re.sub(r'\n\s*-\s*\n', '\n', text)    # empty list items
+        text = re.sub(r'  +', ' ', text)              # double spaces
+        text = re.sub(r'\n{3,}', '\n\n', text)        # triple+ newlines
+
+        return text.strip()
+
     def start_conversation(
         self,
         user_id: int,
@@ -294,6 +349,9 @@ class NordaBizChatEngine:
                 user_context=user_context
             )
 
+            # CRITICAL: Validate all company references — remove hallucinated firms
+            response = self._validate_company_references(response)
+
             # Calculate metrics for per-message tracking in AIChatMessage table
             latency_ms = int((time.time() - start_time) * 1000)
             if self.gemini_service:
@@ -1644,6 +1702,9 @@ W dyskusji [Artur Wiertel](link) pytał o moderację. Pełna treść: [moje uwag
             # Post-process links in full response
             full_response_text = self._postprocess_links(full_response_text, context)
 
+            # CRITICAL: Validate all company references — remove hallucinated firms
+            full_response_text = self._validate_company_references(full_response_text)
+
             # Calculate metrics
             latency_ms = int((time.time() - start_time) * 1000)
             input_tokens = len(full_prompt) // 4