Update app.py

app.py

@@ -1,59 +1,76 @@
+from huggingface_hub import snapshot_download
+import os
+
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+print("Downloading MuRIL model...")
+muril_path = snapshot_download(
+    repo_id="nitz0219/bargainai-muril",
+    repo_type="model",
+    token=HF_TOKEN
+)
+print("MuRIL ready!")
+
+print("Downloading BERT model...")
+bert_path = snapshot_download(
+    repo_id="nitz0219/bargainai-bert",
+    repo_type="model",
+    token=HF_TOKEN
+)
+print("BERT ready!")
+
+index_path = os.path.join(muril_path, "index")
+print(f"Index path: {index_path}")
+print(f"Index files: {os.listdir(index_path)}")
+
 import gradio as gr
 import json
 import numpy as np
 import torch
-import os
 import anthropic
 from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
 
-MURIL_MODEL = "models/muril_finetuned"
-BERT_MODEL = "models/bert_finetuned"
-
-
 class BargainingAgent:
     def __init__(self):
-        self.device = torch.device(
-            "mps" if torch.backends.mps.is_available() else "cpu")
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        print(f"Device: {self.device}")
 
         print("Loading India model (MuRIL)...")
-        self.muril_embeddings = np.load("index/muril_finetuned_embeddings.npy")
-        with open("index/muril_finetuned_metadata.json") as f:
+        self.muril_embeddings = np.load(os.path.join(index_path, "muril_finetuned_embeddings.npy"))
+        with open(os.path.join(index_path, "muril_finetuned_metadata.json")) as f:
             self.muril_metadata = json.load(f)
-        with open("models/muril_finetuned/label_map.json") as f:
+        with open(os.path.join(muril_path, "label_map.json")) as f:
             lm = json.load(f)
         self.muril_id_to_label = {v: k for k, v in lm.items()}
-        self.muril_tokenizer = AutoTokenizer.from_pretrained(MURIL_MODEL)
-        self.muril_classifier = AutoModelForSequenceClassification.from_pretrained(
-            MURIL_MODEL).to(self.device)
+        self.muril_tokenizer = AutoTokenizer.from_pretrained(muril_path)
+        self.muril_classifier = AutoModelForSequenceClassification.from_pretrained(muril_path).to(self.device)
         self.muril_classifier.eval()
-        self.muril_embed = AutoModel.from_pretrained(
-            MURIL_MODEL).to(self.device)
+        self.muril_embed = AutoModel.from_pretrained(muril_path).to(self.device)
         self.muril_embed.eval()
+        print("MuRIL loaded!")
 
         print("Loading Global model (BERT)...")
-        self.bert_embeddings = np.load("index/bert_finetuned_embeddings.npy")
-        with open("index/bert_finetuned_metadata.json") as f:
+        self.bert_embeddings = np.load(os.path.join(index_path, "bert_finetuned_embeddings.npy"))
+        with open(os.path.join(index_path, "bert_finetuned_metadata.json")) as f:
             self.bert_metadata = json.load(f)
-        with open("models/bert_finetuned/label_map.json") as f:
+        with open(os.path.join(bert_path, "label_map.json")) as f:
             lm2 = json.load(f)
         self.bert_id_to_label = {v: k for k, v in lm2.items()}
-        self.bert_tokenizer = AutoTokenizer.from_pretrained(BERT_MODEL)
-        self.bert_classifier = AutoModelForSequenceClassification.from_pretrained(
-            BERT_MODEL).to(self.device)
+        self.bert_tokenizer = AutoTokenizer.from_pretrained(bert_path)
+        self.bert_classifier = AutoModelForSequenceClassification.from_pretrained(bert_path).to(self.device)
         self.bert_classifier.eval()
-        self.bert_embed = AutoModel.from_pretrained(BERT_MODEL).to(self.device)
+        self.bert_embed = AutoModel.from_pretrained(bert_path).to(self.device)
         self.bert_embed.eval()
+        print("BERT loaded!")
 
-        self.claude = anthropic.Anthropic(
-            api_key=os.environ.get("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"))
+        self.claude = anthropic.Anthropic(api_key=os.environ.get("ANTHROPIC_API_KEY"))
         print("Both models ready!")
 
     def classify_intent(self, text, market):
         tokenizer = self.muril_tokenizer if market == "india" else self.bert_tokenizer
         classifier = self.muril_classifier if market == "india" else self.bert_classifier
         id_to_label = self.muril_id_to_label if market == "india" else self.bert_id_to_label
-        enc = tokenizer(text, truncation=True, padding=True,
-                        max_length=128, return_tensors="pt").to(self.device)
+        enc = tokenizer(text, truncation=True, padding=True, max_length=128, return_tensors="pt").to(self.device)
         with torch.no_grad():
             out = classifier(**enc)
         probs = torch.softmax(out.logits, dim=1)
@@ -63,8 +80,7 @@ class BargainingAgent:
     def get_embedding(self, text, market):
         tokenizer = self.muril_tokenizer if market == "india" else self.bert_tokenizer
         embed_model = self.muril_embed if market == "india" else self.bert_embed
-        enc = tokenizer(text, truncation=True, padding=True,
-                        max_length=128, return_tensors="pt").to(self.device)
+        enc = tokenizer(text, truncation=True, padding=True, max_length=128, return_tensors="pt").to(self.device)
         with torch.no_grad():
             out = embed_model(**enc)
         emb = out[0].mean(dim=1)
@@ -76,8 +92,7 @@ class BargainingAgent:
         qe = self.get_embedding(query, market)
         scores = np.dot(embeddings, qe.T).flatten()
         if pillar:
-            filtered = [i for i, m in enumerate(
-                metadata) if m["pillar"].lower() == pillar.lower()]
+            filtered = [i for i, m in enumerate(metadata) if m["pillar"].lower() == pillar.lower()]
             fs = np.full(len(embeddings), -1.0)
             for i in filtered:
                 fs[i] = scores[i]
@@ -101,8 +116,7 @@ class BargainingAgent:
         }.get(intent, "Negotiation")
 
     def next_price(self, intent, current, floor):
-        drops = {"price_objection": 20,
-                 "walkaway_threat": 30, "ready_to_buy": 0}
+        drops = {"price_objection": 20, "walkaway_threat": 30, "ready_to_buy": 0}
         drop = drops.get(intent, 10)
         return max(current - drop, floor)
 
@@ -116,38 +130,29 @@ class BargainingAgent:
         if market == "india":
             if gender == "Female":
                 address = "Didi"
-                tone = "warm, sisterly, lightly playful — like a friendly older sister who runs a shop"
-                example = "Arre Didi, aapki choice ekdum amazing hai! Aapke liye Rs.820 final kar deti hoon — free delivery bhi milegi 😊"
+                tone = "warm, sisterly, lightly playful"
+                example = "Arre Didi, aapki choice ekdum amazing hai! Rs.820 final kar deti hoon 😊"
             elif gender == "Male":
                 address = "Bhaiya"
-                tone = "friendly, brotherly, light humor — like a fun shopkeeper who enjoys chatting"
-                example = "Bhaiya aapki nazar toh ekdum sahi jagah padi! Rs.820 mein le jao — best deal hai aaj ka, guarantee se 😄"
+                tone = "friendly, brotherly, light humor"
+                example = "Bhaiya aapki nazar sahi jagah padi! Rs.820 mein le jao 😄"
             else:
                 address = "Aap"
                 tone = "warm, respectful, light humor"
-                example = "Aapke liye Rs.820 final kar deta hoon — ekdum best price hai yeh aaj ke liye 😊"
+                example = "Aapke liye Rs.820 final kar deta hoon 😊"
 
             language_instruction = f"""Respond in natural Hinglish (Hindi + English mix).
-Address the customer as {address}.
-Tone: {tone}
-
-Important rules:
-- Add light humor or a playful comment naturally — not forced
-- Praise the customer's choice genuinely — make them feel good
-- Use warmth — make them feel special, not just another customer
-- Never sound desperate or robotic or boring
-- Keep it 2-3 sentences maximum
+Address customer as {address}. Tone: {tone}
+- Add light humor naturally
+- Praise customer choice genuinely
+- 2-3 sentences max
 - 1-2 emojis only
-- Use scarcity or urgency naturally if relevant
-
-Good example: '{example}'"""
+Example: '{example}'"""
 
         else:
             language_instruction = """Respond in natural conversational English.
-Warm, confident, friendly and slightly playful — like a real shop owner who genuinely enjoys helping customers.
-Add a light compliment about their choice. Make them feel like they got a great deal.
-Never sound like a bot. Keep it short — 2-3 sentences max.
-Good example: 'Great taste! I can do Rs.820 for you — honestly that is our best price and you are getting a steal today 😄'"""
+Warm, friendly, slightly playful. Add light compliment.
+2-3 sentences max. Example: 'Great taste! Rs.820 is our best price today 😄'"""
 
         prompt = f"""You are an experienced shopkeeper negotiating on WhatsApp for {product_name}.
 
@@ -157,13 +162,12 @@ Customer intent: {intent}
 Negotiation tactic from {book} Page {page}:
 {tactic}
 
-Price: Previous Rs.{current_offer} — New offer Rs.{next_offer}
-{'This is your FINAL price. Do not go lower under any circumstances.' if at_floor else 'You can negotiate slightly more if genuinely needed.'}
+Price: Previous Rs.{current_offer} to New offer Rs.{next_offer}
+{'This is FINAL price. Do not go lower.' if at_floor else 'Can negotiate slightly more if needed.'}
 
-Instructions:
 {language_instruction}
 
-Write only the WhatsApp reply message. Nothing else."""
+Write only the WhatsApp reply."""
 
         message = self.claude.messages.create(
             model="claude-haiku-4-5-20251001",
@@ -171,7 +175,7 @@ Write only the WhatsApp reply message. Nothing else."""
             messages=[{"role": "user", "content": prompt}]
         )
         response = message.content[0].text.strip()
-        info = f"Intent: {intent} ({conf:.0%}) | Source: {book[:35]}... Page {page} | Offer: Rs.{current_offer} → Rs.{next_offer} | Gender: {gender}"
+        info = f"Intent: {intent} ({conf:.0%}) | Source: {book[:35]}... Page {page} | Offer: Rs.{current_offer} to Rs.{next_offer}"
         history.append({"role": "user", "content": customer_msg})
         history.append({"role": "assistant", "content": response})
         return history, info, next_offer, ""
@@ -183,13 +187,11 @@ print("Agent ready!")
 
 CSS = """
 @import url('https://fonts.googleapis.com/css2?family=DM+Sans:wght@300;400;500&display=swap');
-
 body, .gradio-container {
     font-family: 'DM Sans', sans-serif !important;
     background: #0a1628 !important;
     color: #e2e8f0 !important;
 }
-
 .stack-info {
     background: #0f1f35;
     border: 0.5px solid #1e3a5f;
@@ -200,7 +202,6 @@ body, .gradio-container {
     color: #475569;
     line-height: 2;
 }
-
 .footer-note {
     text-align: center;
     color: #1e3a5f;
@@ -211,7 +212,6 @@ body, .gradio-container {
 }
 """
 
-
 def chat(message, history, current_offer, floor_price, mrp, product_name, market, gender):
     if not message.strip():
         return history, "", current_offer, ""
@@ -222,16 +222,13 @@ def chat(message, history, current_offer, floor_price, mrp, product_name, market
     )
     return history, info, new_offer, ""
 
-
 def reset_chat(mrp):
     starting = int(float(mrp) * 0.94)
     return [], "Configure your product and start negotiating...", starting
 
-
 def update_price(x):
     return x
 
-
 def set_q1(): return "bahut mehnga hai bhaiya"
 def set_q2(): return "quality acchi nahi lagti"
 def set_q3(): return "Amazon pe sasta milega"
@@ -241,8 +238,7 @@ def set_q6(): return "I can find it cheaper elsewhere"
 def set_q7(): return "Can you do better on price?"
 def set_q8(): return "Ok I will take it"
 
-
-with gr.Blocks(title="BargainAI") as demo:
+with gr.Blocks(title="BargainAI", css=CSS) as demo:
 
     gr.HTML("""
     <div style="text-align:center;padding:2rem 1rem 1.5rem;background:linear-gradient(180deg,#0f2744 0%,#0a1628 100%);border-bottom:0.5px solid #1e3a5f;margin-bottom:1rem;">
@@ -256,10 +252,8 @@ with gr.Blocks(title="BargainAI") as demo:
     """)
 
     with gr.Row():
-
         with gr.Column(scale=1):
             gr.HTML("<div style='color:#60a5fa;font-size:0.82rem;font-weight:600;letter-spacing:1px;margin-bottom:8px;'>PRODUCT CONFIG</div>")
-
             market = gr.Radio(
                 choices=["India (Hinglish)", "Global (English)"],
                 value="India (Hinglish)",
@@ -270,14 +264,11 @@ with gr.Blocks(title="BargainAI") as demo:
                 value="Male",
                 label="Customer Gender"
             )
-            product_name = gr.Textbox(
-                value="Premium Cotton Kurti", label="Product Name")
+            product_name = gr.Textbox(value="Premium Cotton Kurti", label="Product Name")
             mrp = gr.Number(value=899, label="MRP (Rs.)")
-            floor_price = gr.Number(
-                value=749, label="Floor Price — Never go below")
+            floor_price = gr.Number(value=749, label="Floor Price — Never go below")
             current_offer = gr.State(value=849)
-            price_display = gr.Number(
-                value=849, label="Live Offer Price (Rs.)", interactive=False)
+            price_display = gr.Number(value=849, label="Live Offer Price (Rs.)", interactive=False)
             reset_btn = gr.Button("Reset Conversation", variant="secondary")
 
             gr.HTML("""
@@ -286,7 +277,6 @@ with gr.Blocks(title="BargainAI") as demo:
                 India: MuRIL fine-tuned · 80.5% accuracy<br>
                 Global: BERT fine-tuned · 83.17% accuracy<br>
                 Gender-aware · Bhaiya / Didi / Aap<br>
-                Warm humor · Natural Hinglish<br>
                 9 books · 2,383 indexed chunks<br>
                 Claude Haiku · Real-time responses
             </div>
@@ -307,16 +297,14 @@ with gr.Blocks(title="BargainAI") as demo:
                 interactive=False
             )
 
-            gr.HTML(
-                "<div style='color:#475569;font-size:0.78rem;margin:8px 0 4px;'>India quick replies:</div>")
+            gr.HTML("<div style='color:#475569;font-size:0.78rem;margin:8px 0 4px;'>India quick replies:</div>")
             with gr.Row():
                 q1 = gr.Button("bahut mehnga hai", size="sm")
                 q2 = gr.Button("quality acchi nahi", size="sm")
                 q3 = gr.Button("Amazon pe sasta", size="sm")
                 q4 = gr.Button("final price kya hai", size="sm")
 
-            gr.HTML(
-                "<div style='color:#475569;font-size:0.78rem;margin:8px 0 4px;'>Global quick replies:</div>")
+            gr.HTML("<div style='color:#475569;font-size:0.78rem;margin:8px 0 4px;'>Global quick replies:</div>")
             with gr.Row():
                 q5 = gr.Button("Too expensive", size="sm")
                 q6 = gr.Button("Cheaper elsewhere", size="sm")
@@ -334,7 +322,7 @@ with gr.Blocks(title="BargainAI") as demo:
 
     gr.HTML("""
     <div class="footer-note">
-        Built by Nitesh Nankani · MuRIL + BERT + Claude Haiku + 9 Negotiation Books · Python · HuggingFace · Gradio
+        Built by Nitesh Nankani · MuRIL + BERT + Claude Haiku + 9 Negotiation Books �� HuggingFace · Gradio
     </div>
     """)
 
@@ -349,15 +337,13 @@ with gr.Blocks(title="BargainAI") as demo:
 
     send_btn.click(
         fn=chat,
-        inputs=[msg_input, chatbot, current_offer,
-                floor_price, mrp, product_name, market, gender],
+        inputs=[msg_input, chatbot, current_offer, floor_price, mrp, product_name, market, gender],
         outputs=[chatbot, intel_bar, current_offer, msg_input]
     ).then(fn=update_price, inputs=[current_offer], outputs=[price_display])
 
     msg_input.submit(
         fn=chat,
-        inputs=[msg_input, chatbot, current_offer,
-                floor_price, mrp, product_name, market, gender],
+        inputs=[msg_input, chatbot, current_offer, floor_price, mrp, product_name, market, gender],
         outputs=[chatbot, intel_bar, current_offer, msg_input]
     ).then(fn=update_price, inputs=[current_offer], outputs=[price_display])
 
@@ -368,4 +354,4 @@ with gr.Blocks(title="BargainAI") as demo:
     ).then(fn=update_price, inputs=[current_offer], outputs=[price_display])
 
 if __name__ == "__main__":
-    demo.launch(share=True)
+    demo.launch()