Evidence-grounded LLM verdicts (date context, think-block-safe JSON parsing, source status)

src/streamlit_app.py

@@ -1,4 +1,5 @@
 import os
+import re
 import json
 import time
 import requests
@@ -452,6 +453,8 @@ def parse_llm_json(raw):
     if not raw:
         return None
     text = raw.strip()
+    if "<think>" in text:
+        text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL).strip()
     if text.startswith("```"):
         text = text.strip("`")
         if text.startswith("json"):
@@ -475,6 +478,12 @@ SYSTEM_PROMPT = (
     "app. You will receive a piece of content, two AI classifier verdicts with confidence "
     "scores, and web search evidence. Your job is to decide how truthful the content is. "
     "You are the final judge: your verdict overrides the classifier verdicts.\n"
+    "CRITICAL: The web evidence you are given is the ground truth. Base your verdict "
+    "primarily on it: if the evidence supports the claim, choose REAL; if it contradicts "
+    "the claim, choose FAKE; if no evidence was provided or it is genuinely insufficient, "
+    "choose UNKNOWN. You must never claim that evidence does not exist online — you can "
+    "only judge the evidence you were provided.\n"
+    "Pay attention to the current date you are given: events in the past are not future events.\n"
     'Respond ONLY with a valid JSON object in this exact shape:\n'
     '{"verdict": "REAL" or "FAKE" or "UNKNOWN", "confidence": <integer 0-100>, '
     '"explanation": "<2-3 plain-language sentences>", '
@@ -492,13 +501,19 @@ def synthesize_verdict(content, model_results, evidence):
     if not GROQ_AVAILABLE:
         return None
     payload = {
+        "current_date": datetime.now().strftime("%B %d, %Y"),
         "content": content,
         "classifier_verdicts": {k: v for k, v in model_results.items() if v},
         "web_evidence": [
-            {"title": s["title"], "content": s["content"][:200]} for s in evidence
+            {
+                "title": s["title"],
+                "url": s["url"],
+                "content": s["content"][:250],
+            }
+            for s in evidence
         ],
     }
-    prompt = json.dumps(payload, ensure_ascii=False)[:6000]
+    prompt = json.dumps(payload, ensure_ascii=False)[:7000]
     for model in GROQ_MODELS:
         try:
             response = GROQ_CLIENT.chat.completions.create(
@@ -508,7 +523,7 @@ def synthesize_verdict(content, model_results, evidence):
                     {"role": "user", "content": prompt},
                 ],
                 temperature=0.2,
-                max_tokens=700,
+                max_tokens=1100,
                 response_format={"type": "json_object"},
             )
             data = parse_llm_json(response.choices[0].message.content)
@@ -628,6 +643,10 @@ def process_analysis(user_input, input_method, clf_political, clf_general):
 
         st.write("🌐 Checking claims across the web...")
         result = run_analysis(text_to_analyze, clf_political, clf_general)
+        if result.get("evidence"):
+            st.write(f"🌐 Found {len(result['evidence'])} relevant sources")
+        else:
+            st.write("ℹ️ No web results found — analysis based on AI review")
 
         analysis_time = time.time() - start_time
         status.update(label="✅ Analysis ready!", state="complete")