improved chat

app.py

@@ -103,16 +103,10 @@ PREAMBLE_RES = [
 ]
 
 OPTIMIZER_PROMPT = (
-    "You are a search query optimizer. Your job is to translate user questions into database search keywords.\n"
+    "You are a search query optimizer. Your job is to translate user questions into database search keywords. "
     "Translate synonyms into our core product areas: 'onboarding/acquisition' -> 'DigiOnboard', "
-    "'payments' -> 'SmartPayments', 'voice/soundbox' -> 'SoundHub', 'cross-sell' -> 'SellSmart'.\n"
-    "Output ONLY the keywords, separated by spaces. Do not write explanations or punctuation.\n\n"
-    "Examples:\n"
-    "User: Tell me about Mintoak's loyalty program -> Keywords: loyalty program\n"
-    "User: How to onboard a merchant? -> Keywords: DigiOnboard merchant onboarding\n"
-    "User: What is the soundbox device? -> Keywords: SoundHub soundbox device\n"
-    "User: How do they accept payments? -> Keywords: SmartPayments accept payments\n"
-    "User: Tell me about business360 -> Keywords: business360"
+    "'payments' -> 'SmartPayments', 'voice/soundbox' -> 'SoundHub', 'cross-sell' -> 'SellSmart'. "
+    "Output ONLY the keywords, separated by spaces. Do not write explanations or punctuation."
 )
 
 SYSTEM_PROMPT = (
@@ -148,7 +142,8 @@ QUERY_ENHANCEMENT_RULES = [
     },
     {
         "keywords": ["loyalty", "reward", "rewards", "campaign", "rewardrun"],
-        "expansion": "Mintoak RewardRun gamified loyalty campaigns merchant activation retention",
+        "expansion": "loyalty_programs",
+        "override": True,
         "prompt_note": "Ensure you mention Mintoak RewardRun as the gamified loyalty and merchant engagement platform.",
         "top_k": 3
     }
@@ -156,6 +151,31 @@ QUERY_ENHANCEMENT_RULES = [
 
 MAX_CHUNK_CHARS = 500
 
+def has_lead_intent(query: str) -> bool:
+    q = query.lower()
+    # Check for demo requests
+    if any(kw in q for kw in ["book a demo", "schedule a demo", "request a demo", "book demo", "schedule demo", "request demo", "get a demo", "live demo"]):
+        return True
+    # Check for pricing/cost queries
+    if any(kw in q for kw in ["pricing", "price list", "pricing plans", "subscription cost", "licensing cost", "what are the charges", "setup fees"]):
+        return True
+    if "cost" in q and any(w in q for w in ["how much", "what is", "what's"]):
+        return True
+    if "how much" in q and any(w in q for w in ["cost", "charge", "pay", "fee"]):
+        return True
+    # Check for partnership queries
+    if any(kw in q for kw in ["become a partner", "partnering with you", "partnership opportunities", "partner with mintoak"]):
+        return True
+    # Check for contact/sales queries
+    if any(kw in q for kw in ["talk to sales", "sales team", "contact sales", "sales department", "sales contact"]):
+        return True
+    if any(kw in q for kw in ["get in touch", "how to contact", "contact us", "contact details", "callback", "call back"]):
+        return True
+    # Check for request to contact
+    if "call me" in q or "phone number" in q or "email address" in q:
+        return True
+    return False
+
 def retrieve_context(query: str, top_k: int = 2):
     query_lower = query.lower()
     search_keywords = None
@@ -164,7 +184,10 @@ def retrieve_context(query: str, top_k: int = 2):
     # Step 1: Check config-driven expansion registry first
     for rule in QUERY_ENHANCEMENT_RULES:
         if any(kw in query_lower for kw in rule["keywords"]):
-            search_keywords = f"{query} {rule['expansion']}"
+            if rule.get("override"):
+                search_keywords = rule["expansion"]
+            else:
+                search_keywords = f"{query} {rule['expansion']}"
             top_k = rule["top_k"]
             prompt_note = rule.get("prompt_note")
             break
@@ -173,6 +196,16 @@ def retrieve_context(query: str, top_k: int = 2):
     if not search_keywords:
         messages = [
             {"role": "system", "content": OPTIMIZER_PROMPT},
+            {"role": "user", "content": "Tell me about Mintoak's loyalty program"},
+            {"role": "assistant", "content": "loyalty program"},
+            {"role": "user", "content": "How to onboard a merchant?"},
+            {"role": "assistant", "content": "DigiOnboard merchant onboarding"},
+            {"role": "user", "content": "What is the soundbox device?"},
+            {"role": "assistant", "content": "SoundHub soundbox device"},
+            {"role": "user", "content": "How do they accept payments?"},
+            {"role": "assistant", "content": "SmartPayments accept payments"},
+            {"role": "user", "content": "Tell me about business360"},
+            {"role": "assistant", "content": "business360"},
             {"role": "user",   "content": query}
         ]
         prompt = _tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
@@ -231,9 +264,34 @@ def retrieve_context(query: str, top_k: int = 2):
 
     return "\n\n---\n\n".join(context_parts), url_to_title, prompt_note
 
-def clean_response(text: str) -> str:
+def clean_response(text: str, allow_lead_capture: bool = True) -> str:
     for rx in PREAMBLE_RES:
         text = rx.sub("", text).strip()
+    
+    # Strip any model-generated citation lines
+    text = re.sub(r'(?i)(?:\n\s*)*👉?\s*For\s+more\s+details,?\s+visit:.*$', '', text).strip()
+    
+    has_lead = "[CAPTURE_LEAD]" in text and allow_lead_capture
+    text_no_lead = text.replace("[CAPTURE_LEAD]", "").strip()
+    text_lower = text_no_lead.lower()
+    
+    fallback_triggers = [
+        "not fully answered", "not mentioned", "not provided", "unable to answer", 
+        "cannot be determined", "no mention", "not find any information", 
+        "does not contain", "no information", "not in the context", "not explicitly mentioned",
+        "isn't mentioned", "not found"
+    ]
+    if any(trigger in text_lower for trigger in fallback_triggers):
+        text_no_lead = (
+            "The requested information does not currently exist on www.mintoak.com. "
+            "You can get in touch with our team at https://www.mintoak.com/contact-us."
+        )
+
+    if has_lead:
+        text = text_no_lead + " [CAPTURE_LEAD]"
+    else:
+        text = text_no_lead
+
     if text and text[0].islower():
         text = text[0].upper() + text[1:]
     return text
@@ -330,6 +388,7 @@ def index():
 def chat():
     data = request.get_json(force=True)
     query = (data.get("query") or "").strip()
+    history = data.get("history") or []
     if not query:
         return Response("data: " + json.dumps({"error": "Empty query"}) + "\n\n", mimetype="text/event-stream")
 
@@ -372,14 +431,22 @@ def chat():
     def stream_rag():
         yield f"data: {json.dumps({'event': 'sources', 'sources': sources})}\n\n"
 
+        num_user_questions = sum(1 for msg in history if msg.get("role") == "user") + 1
+        allow_lead_capture = has_lead_intent(query) or (num_user_questions >= 3)
+
         user_content = f"Context:\n{context}\n\nQuestion: {query}"
         if prompt_note:
             user_content += f"\n\nNote: {prompt_note}"
 
-        messages = [
-            {"role": "system", "content": SYSTEM_PROMPT},
-            {"role": "user",   "content": user_content}
-        ]
+        # Build messages from system prompt + conversation history + current user query
+        messages = [{"role": "system", "content": SYSTEM_PROMPT}]
+        for msg in history:
+            role = "user" if msg.get("role") == "user" else "assistant"
+            content = msg.get("content", "").strip()
+            content = content.replace("[CAPTURE_LEAD]", "").strip()
+            messages.append({"role": role, "content": content})
+        
+        messages.append({"role": "user",   "content": user_content})
         prompt = _tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
 
         t0 = time.time()
@@ -403,21 +470,36 @@ def chat():
             thread.start()
             
             citation_started = False
+            lead_tag_sent = False
             for token_text in streamer:
                 full_response += token_text
                 
                 if citation_started:
+                    if "[CAPTURE_LEAD]" in full_response and not lead_tag_sent and allow_lead_capture:
+                        lead_tag_sent = True
+                        yield f"data: {json.dumps({'event': 'token', 'token': '[CAPTURE_LEAD]'})}\n\n"
                     continue
                 
                 last_snippet = full_response[-40:].lower()
                 if "👉" in token_text or "for more details" in last_snippet or "details, visit" in last_snippet:
                     citation_started = True
                     continue
+                
+                if "[CAPTURE_LEAD]" in token_text:
+                    if allow_lead_capture and not lead_tag_sent:
+                        lead_tag_sent = True
+                        yield f"data: {json.dumps({'event': 'token', 'token': '[CAPTURE_LEAD]'})}\n\n"
+                    continue
                     
                 yield f"data: {json.dumps({'event': 'token', 'token': token_text})}\n\n"
+            
+            # Ensure lead tag is sent if generated or proactively triggered
+            should_send_lead = allow_lead_capture and ("[CAPTURE_LEAD]" in full_response or num_user_questions >= 3)
+            if should_send_lead and not lead_tag_sent:
+                yield f"data: {json.dumps({'event': 'token', 'token': ' [CAPTURE_LEAD]'})}\n\n"
 
         latency = round(time.time() - t0, 2)
-        cleaned_response = clean_response(full_response.strip())
+        cleaned_response = clean_response(full_response.strip(), allow_lead_capture=allow_lead_capture)
 
         # If references were not generated, append them
         if "Source:" not in cleaned_response and sources:

scripts/mintoak/chat_server.py

@@ -119,7 +119,7 @@ QUERY_ENHANCEMENT_RULES = [
     },
     {
         "keywords": ["document", "documents", "checklist", "paperwork", "certificates"],
-        "expansion": "Mintoak DigiOnboard merchant onboarding documents checklist GST PAN registration certificates cancelled cheques board resolutions",
+        "expansion": "What documents are needed for merchant onboarding? DigiOnboard KYC KYB",
         "prompt_note": "Ensure you list the specific documents needed: registration certificates, GST certificates, cancelled cheques, and board resolutions.",
         "top_k": 3
     },
@@ -128,11 +128,43 @@ QUERY_ENHANCEMENT_RULES = [
         "expansion": "Mintoak DigiOnboard merchant onboarding KYC KYB acquisition banks",
         "prompt_note": None,
         "top_k": 2
+    },
+    {
+        "keywords": ["loyalty", "reward", "rewards", "campaign", "rewardrun"],
+        "expansion": "loyalty_programs",
+        "override": True,
+        "prompt_note": "Ensure you mention Mintoak RewardRun as the gamified loyalty and merchant engagement platform.",
+        "top_k": 3
     }
 ]
 
 MAX_CHUNK_CHARS = 400   # Truncate each chunk to cap prompt size & speed up prefill
 
+def has_lead_intent(query: str) -> bool:
+    q = query.lower()
+    # Check for demo requests
+    if any(kw in q for kw in ["book a demo", "schedule a demo", "request a demo", "book demo", "schedule demo", "request demo", "get a demo", "live demo"]):
+        return True
+    # Check for pricing/cost queries
+    if any(kw in q for kw in ["pricing", "price list", "pricing plans", "subscription cost", "licensing cost", "what are the charges", "setup fees"]):
+        return True
+    if "cost" in q and any(w in q for w in ["how much", "what is", "what's"]):
+        return True
+    if "how much" in q and any(w in q for w in ["cost", "charge", "pay", "fee"]):
+        return True
+    # Check for partnership queries
+    if any(kw in q for kw in ["become a partner", "partnering with you", "partnership opportunities", "partner with mintoak"]):
+        return True
+    # Check for contact/sales queries
+    if any(kw in q for kw in ["talk to sales", "sales team", "contact sales", "sales department", "sales contact"]):
+        return True
+    if any(kw in q for kw in ["get in touch", "how to contact", "contact us", "contact details", "callback", "call back"]):
+        return True
+    # Check for request to contact
+    if "call me" in q or "phone number" in q or "email address" in q:
+        return True
+    return False
+
 def retrieve_context(query: str, history: list = None, top_k: int = 2):
     """Return (context_text, url_to_title_dict, prompt_note) or ('OUT_OF_SCOPE_REFUSAL', {}, None)."""
     query_lower = query.lower()
@@ -166,7 +198,10 @@ def retrieve_context(query: str, history: list = None, top_k: int = 2):
     # Apply configuration-driven expansion rules dynamically (without wiping out original query keywords)
     for rule in QUERY_ENHANCEMENT_RULES:
         if any(kw in query_lower for kw in rule["keywords"]):
-            search_query = f"{query} {rule['expansion']}"
+            if rule.get("override"):
+                search_query = rule["expansion"]
+            else:
+                search_query = f"{query} {rule['expansion']}"
             top_k = rule["top_k"]
             prompt_note = rule.get("prompt_note")
             break
@@ -216,7 +251,7 @@ def retrieve_context(query: str, history: list = None, top_k: int = 2):
     return "\n\n---\n\n".join(context_parts), url_to_title, prompt_note
 
 
-def clean_response(text: str) -> str:
+def clean_response(text: str, allow_lead_capture: bool = True) -> str:
     """Strip robotic preambles and fix capitalisation."""
     for rx in PREAMBLE_RES:
         text = rx.sub("", text).strip()
@@ -225,7 +260,7 @@ def clean_response(text: str) -> str:
     text = re.sub(r'(?i)(?:\n\s*)*👉?\s*For\s+more\s+details,?\s+visit:.*$', '', text).strip()
     
     # Handle LLM fallback non-compliance phrases and override with the brand-approved fallback
-    has_lead = "[CAPTURE_LEAD]" in text
+    has_lead = "[CAPTURE_LEAD]" in text and allow_lead_capture
     text_no_lead = text.replace("[CAPTURE_LEAD]", "").strip()
     text_lower = text_no_lead.lower()
     
@@ -432,6 +467,9 @@ def chat():
         # First send sources so UI can render them instantly
         yield f"data: {json.dumps({'event': 'sources', 'sources': sources})}\n\n"
 
+        num_user_questions = sum(1 for msg in history if msg.get("role") == "user") + 1
+        allow_lead_capture = has_lead_intent(query) or (num_user_questions >= 3)
+
         # Rewrite the query passed to the LLM if it is a conversational feedback query
         llm_query = query
         if history:
@@ -478,7 +516,7 @@ def chat():
                 full_response += token_text
                 
                 if citation_started:
-                    if "[CAPTURE_LEAD]" in full_response and not lead_tag_sent:
+                    if "[CAPTURE_LEAD]" in full_response and not lead_tag_sent and allow_lead_capture:
                         lead_tag_sent = True
                         yield f"data: {json.dumps({'event': 'token', 'token': '[CAPTURE_LEAD]'})}\n\n"
                     continue
@@ -488,14 +526,21 @@ def chat():
                     citation_started = True
                     continue
                 
+                if "[CAPTURE_LEAD]" in token_text:
+                    if allow_lead_capture and not lead_tag_sent:
+                        lead_tag_sent = True
+                        yield f"data: {json.dumps({'event': 'token', 'token': '[CAPTURE_LEAD]'})}\n\n"
+                    continue
+                
                 yield f"data: {json.dumps({'event': 'token', 'token': token_text})}\n\n"
             
-            # Ensure lead tag is sent if generated but not yet streamed
-            if "[CAPTURE_LEAD]" in full_response and not lead_tag_sent:
-                yield f"data: {json.dumps({'event': 'token', 'token': '[CAPTURE_LEAD]'})}\n\n"
+            # Ensure lead tag is sent if generated or proactively triggered
+            should_send_lead = allow_lead_capture and ("[CAPTURE_LEAD]" in full_response or num_user_questions >= 3)
+            if should_send_lead and not lead_tag_sent:
+                yield f"data: {json.dumps({'event': 'token', 'token': ' [CAPTURE_LEAD]'})}\n\n"
 
         latency = round(time.time() - t0, 2)
-        cleaned_response = clean_response(full_response.strip())
+        cleaned_response = clean_response(full_response.strip(), allow_lead_capture=allow_lead_capture)
 
         # If references were not generated in the response body, append them dynamically
         if "Source:" not in cleaned_response and "details, visit:" not in cleaned_response and sources: