Update llm_client.py

llm_client.py

@@ -120,7 +120,8 @@ class GroqLLM(LLM):
                 model=self.groq_model,
                 temperature=0.3,
                 max_tokens=1024,
-                stop=stop_seq
+                stop=stop_seq,
+                extra_body={"enable_thinking": True}
             )
             return chat_completion.choices[0].message.content
         except Exception as e:
@@ -204,7 +205,7 @@ class LLMClient:
             self.server_bin,
             "-m", self.model_path,
             "--port", str(self.server_port),
-            "-c", "8192",
+            "-c", "8192", # Increased context window to handle web content
             "--host", "0.0.0.0",
             "--mlock"  # Lock model in RAM to prevent swapping 
         ]
@@ -262,12 +263,13 @@ class LLMClient:
         # Custom Prompt Template
         template = """<|im_start|>system
         You are CyberGuard - an AI specialized in Phishing Detection.
-        Task: Analyze the provided URL and HTML snippet to classify the website as 'PHISHING' or 'BENIGN'.        
+        Task: Analyze the provided URL and HTML snippet to classify the website as 'PHISHING' or 'BENIGN'.
         Check specifically for BRAND IMPERSONATION (e.g. Facebook, Google, Banks).
         If the HTML content is missing, empty, or contains an error message (like "Could not fetch website content"), YOU MUST RETURN classification by ANALYZING the URL.
         Classification Rules:
         - PHISHING: Typosquatting URLs (e.g., paypa1.com), hidden login forms, obfuscated javascript, mismatched branding vs URL.
         - BENIGN: Legitimate website, clean code, URL matches the content/brand.
+        Let's think step by step to verify logical inconsistencies between URL and Content before deciding.
 
         RETURN THE RESULT IN THE EXACT FOLLOWING FORMAT (NO PREAMBLE):