Update app.py

app.py

@@ -16,7 +16,6 @@ OPENROUTER_URL = "https://openrouter.ai/api/v1/chat/completions"
 GEMMA_MODEL = "google/gemma-3-4b-it:free"
 
 st.set_page_config(page_title="EZOFIS Document Validation Agent", layout="wide")
-
 st.markdown("""
     <style>
     .block-card {
@@ -39,7 +38,7 @@ st.markdown(
     unsafe_allow_html=True
 )
 st.markdown(
-    "<div style='font-size:20px; margin-bottom:28px; color:#24345C;'>AI-driven checklist-based document acceptance for mortgage applications.</div>",
+    "<div style='font-size:20px; margin-bottom:28px; color:#24345C;'>AI-driven, agentic document acceptance for mortgage applications.</div>",
     unsafe_allow_html=True
 )
 
@@ -174,6 +173,44 @@ def query_gemma_llm(doc_text, checklist, status_box=None):
             status_box.write(result)
         return None, result, prompt
 
+def advanced_llm_verdict(llm_json, min_confidence, status_box=None):
+    # Only trigger if confidence is in gray zone: [min_confidence, min_confidence+15)
+    conf = llm_json.get("confidence", 0)
+    if conf < min_confidence or conf >= min_confidence + 15:
+        return None, None, None
+    verdict_prompt = f"""
+Here is the extracted document information and prior validation result:
+{json.dumps(llm_json)}
+
+The minimum required confidence is {min_confidence}. Should this document be accepted or rejected for a mortgage application, based on all available information?
+Respond ONLY as: {{ "accepted": true/false, "reason": "..." }}
+"""
+    headers = {
+        "Authorization": f"Bearer {OPENROUTER_API_KEY}",
+        "HTTP-Referer": "https://chat.openai.com",
+        "X-Title": "EZOFIS-Doc-Validator",
+        "Content-Type": "application/json",
+    }
+    data = {
+        "model": GEMMA_MODEL,
+        "messages": [{"role": "user", "content": verdict_prompt}],
+        "temperature": 0.1,
+        "max_tokens": 256
+    }
+    if status_box:
+        status_box.info("Step 3: LLM self-verdict (gray zone confidence)...")
+    resp = requests.post(OPENROUTER_URL, headers=headers, json=data, timeout=60)
+    if resp.status_code == 200:
+        try:
+            content = resp.json()["choices"][0]["message"]["content"]
+            vstart = content.find("{")
+            vend = content.rfind("}") + 1
+            verdict_json = json.loads(content[vstart:vend])
+            return verdict_json, content, verdict_prompt
+        except Exception:
+            return None, content, verdict_prompt
+    return None, None, verdict_prompt
+
 def fuzzy_match_type(detected_type, checklist_types):
     best_type = None
     best_score = 0
@@ -219,6 +256,10 @@ uploaded_files = st.file_uploader(
     accept_multiple_files=True
 )
 
+st.markdown("<span class='step-num'>3</span> <b>Configure Acceptance Thresholds</b>", unsafe_allow_html=True)
+min_match_score = st.slider("Minimum Type Match Score (0-100)", 50, 100, 70, 1)
+min_confidence = st.slider("Minimum LLM Confidence (0-100)", 50, 100, 70, 1)
+
 # ========== PROCESSING ==========
 if st.button("Run Document Validation", type="primary") and uploaded_files:
     results = []
@@ -254,17 +295,17 @@ if st.button("Run Document Validation", type="primary") and uploaded_files:
         detected_type = llm_json.get("document_type", "")
         matched_type, match_score = fuzzy_match_type(detected_type, required_types)
 
-        # Accept only if LLM states checklist_matched, looks genuine, and not expired
+        # Accept only if LLM states checklist_matched, looks genuine, and not expired, and confidence high enough
         checklist_matched = llm_json.get("checklist_matched", False)
-        if checklist_matched:
-            # Double check: If match_score < 65, override to not matched
-            if match_score < 65:
-                checklist_matched = False
+        if checklist_matched and match_score < min_match_score:
+            checklist_matched = False
 
+        llm_conf = llm_json.get("confidence", 0)
         accepted = (
             checklist_matched and
             llm_json.get("looks_genuine", False) and
-            not llm_json.get("is_expired", False)
+            not llm_json.get("is_expired", False) and
+            (llm_conf >= min_confidence)
         )
 
         reason = []
@@ -279,9 +320,20 @@ if st.button("Run Document Validation", type="primary") and uploaded_files:
             if llm_json.get("is_expired", False):
                 reason.append("Document is expired.")
 
-        reason.append(f"Genuineness confidence: {llm_json.get('confidence', 0)}.")
+        reason.append(f"Genuineness confidence: {llm_conf}.")
         reason.append(llm_json.get("verdict", ""))
 
+        # Advanced agent: If confidence is in a "gray zone", ask the LLM for a final self-verdict
+        verdict_json, verdict_raw, verdict_prompt = advanced_llm_verdict(llm_json, min_confidence, status_box)
+        debug['LLM_self_verdict_prompt'] = verdict_prompt
+        debug['LLM_self_verdict_raw'] = verdict_raw
+        debug['LLM_self_verdict_json'] = verdict_json
+
+        if verdict_json:
+            accepted = verdict_json.get("accepted", False)
+            reason.append(f"LLM Self-verdict: {verdict_json.get('reason','')}")
+            status_box.info("Final decision (gray zone) taken by LLM self-verdict.")
+
         results.append({
             "File": uploaded_file.name,
             "Detected Type": detected_type,
@@ -290,7 +342,7 @@ if st.button("Run Document Validation", type="primary") and uploaded_files:
             "Expiry Date": llm_json.get("expiry_date", "-"),
             "Expired": "Yes" if llm_json.get("is_expired", False) else "No",
             "Genuine": "Yes" if llm_json.get("looks_genuine", False) else "No",
-            "Confidence": llm_json.get("confidence", "-"),
+            "Confidence": llm_conf,
             "Accepted": "Yes" if accepted else "No",
             "Reason": " ".join(reason)
         })