| 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 anthropic |
| from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification |
|
|
| class BargainingAgent: |
| def __init__(self): |
| 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(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(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_path) |
| self.muril_classifier = AutoModelForSequenceClassification.from_pretrained(muril_path).to(self.device) |
| self.muril_classifier.eval() |
| 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(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(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_path) |
| self.bert_classifier = AutoModelForSequenceClassification.from_pretrained(bert_path).to(self.device) |
| self.bert_classifier.eval() |
| 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("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) |
| with torch.no_grad(): |
| out = classifier(**enc) |
| probs = torch.softmax(out.logits, dim=1) |
| prob, idx = torch.max(probs, dim=1) |
| return id_to_label[idx.item()], prob.item() |
|
|
| 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) |
| with torch.no_grad(): |
| out = embed_model(**enc) |
| emb = out[0].mean(dim=1) |
| return torch.nn.functional.normalize(emb, p=2, dim=1).cpu().numpy() |
|
|
| def retrieve_tactic(self, query, market, pillar=None): |
| embeddings = self.muril_embeddings if market == "india" else self.bert_embeddings |
| metadata = self.muril_metadata if market == "india" else self.bert_metadata |
| 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()] |
| fs = np.full(len(embeddings), -1.0) |
| for i in filtered: |
| fs[i] = scores[i] |
| top = np.argsort(fs)[::-1][:2] |
| else: |
| top = np.argsort(scores)[::-1][:2] |
| return metadata[top[0]]["text"][:300], metadata[top[0]]["book"], metadata[top[0]]["page_number"] |
|
|
| def get_pillar(self, intent): |
| return { |
| "price_objection": "Negotiation", |
| "walkaway_threat": "Negotiation", |
| "competitor_comparison": "Sales", |
| "quality_doubt": "Persuasion", |
| "ready_to_buy": "Sales", |
| "guilt_pressure": "Power", |
| "urgent_buyer": "Sales", |
| "value_seeker": "Persuasion", |
| "trust_issue": "Persuasion", |
| "neutral": "Negotiation" |
| }.get(intent, "Negotiation") |
|
|
| def next_price(self, intent, current, floor): |
| drops = {"price_objection": 20, "walkaway_threat": 30, "ready_to_buy": 0} |
| drop = drops.get(intent, 10) |
| return max(current - drop, floor) |
|
|
| def respond(self, customer_msg, current_offer, floor_price, mrp, product_name, market, gender, history): |
| intent, conf = self.classify_intent(customer_msg, market) |
| pillar = self.get_pillar(intent) |
| tactic, book, page = self.retrieve_tactic(customer_msg, market, pillar) |
| next_offer = self.next_price(intent, current_offer, floor_price) |
| at_floor = next_offer == floor_price |
|
|
| if market == "india": |
| if gender == "Female": |
| address = "Didi" |
| 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" |
| 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 😊" |
|
|
| language_instruction = f"""Respond in natural Hinglish (Hindi + English mix). |
| Address customer as {address}. Tone: {tone} |
| - Add light humor naturally |
| - Praise customer choice genuinely |
| - 2-3 sentences max |
| - 1-2 emojis only |
| Example: '{example}'""" |
|
|
| else: |
| language_instruction = """Respond in natural conversational English. |
| 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}. |
| |
| Customer said: "{customer_msg}" |
| Customer intent: {intent} |
| |
| Negotiation tactic from {book} Page {page}: |
| {tactic} |
| |
| 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.'} |
| |
| {language_instruction} |
| |
| Write only the WhatsApp reply.""" |
|
|
| message = self.claude.messages.create( |
| model="claude-haiku-4-5-20251001", |
| max_tokens=180, |
| 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} to Rs.{next_offer}" |
| history.append({"role": "user", "content": customer_msg}) |
| history.append({"role": "assistant", "content": response}) |
| return history, info, next_offer, "" |
|
|
|
|
| print("Loading agent...") |
| agent = BargainingAgent() |
| 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; |
| border-radius: 8px; |
| padding: 1rem; |
| margin-top: 1rem; |
| font-size: 0.78rem; |
| color: #475569; |
| line-height: 2; |
| } |
| .footer-note { |
| text-align: center; |
| color: #1e3a5f; |
| font-size: 0.75rem; |
| padding: 1rem; |
| margin-top: 1rem; |
| border-top: 0.5px solid #1e3a5f; |
| } |
| """ |
|
|
| def chat(message, history, current_offer, floor_price, mrp, product_name, market, gender): |
| if not message.strip(): |
| return history, "", current_offer, "" |
| market_key = "india" if market == "India (Hinglish)" else "global" |
| history, info, new_offer, _ = agent.respond( |
| message, int(current_offer), int(floor_price), int(mrp), |
| product_name, market_key, gender, history |
| ) |
| 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" |
| def set_q4(): return "final price kya hai" |
| def set_q5(): return "This is too expensive" |
| 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", 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;"> |
| <div style="font-size:2.8rem;font-weight:700;background:linear-gradient(135deg,#f5c842,#e8954a);-webkit-background-clip:text;-webkit-text-fill-color:transparent;"> |
| BargainAI |
| </div> |
| <div style="color:#475569;font-size:0.9rem;margin-top:0.3rem;"> |
| AI-powered negotiation agent · India + Global · Gender-aware · 9 books · MuRIL + BERT |
| </div> |
| </div> |
| """) |
|
|
| 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)", |
| label="Market" |
| ) |
| gender = gr.Radio( |
| choices=["Male", "Female", "Unknown"], |
| value="Male", |
| label="Customer Gender" |
| ) |
| 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") |
| current_offer = gr.State(value=849) |
| price_display = gr.Number(value=849, label="Live Offer Price (Rs.)", interactive=False) |
| reset_btn = gr.Button("Reset Conversation", variant="secondary") |
|
|
| gr.HTML(""" |
| <div class="stack-info"> |
| <div style='color:#60a5fa;font-weight:600;margin-bottom:4px;'>Intelligence Stack</div> |
| India: MuRIL fine-tuned · 80.5% accuracy<br> |
| Global: BERT fine-tuned · 83.17% accuracy<br> |
| Gender-aware · Bhaiya / Didi / Aap<br> |
| 9 books · 2,383 indexed chunks<br> |
| Claude Haiku · Real-time responses |
| </div> |
| """) |
|
|
| with gr.Column(scale=2): |
| gr.HTML("<div style='color:#60a5fa;font-size:0.82rem;font-weight:600;letter-spacing:1px;margin-bottom:8px;'>WHATSAPP NEGOTIATION SIMULATOR</div>") |
|
|
| chatbot = gr.Chatbot( |
| value=[], |
| height=420, |
| show_label=False |
| ) |
|
|
| intel_bar = gr.Textbox( |
| value="Configure your product and start negotiating...", |
| label="Agent Intelligence", |
| interactive=False |
| ) |
|
|
| 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>") |
| with gr.Row(): |
| q5 = gr.Button("Too expensive", size="sm") |
| q6 = gr.Button("Cheaper elsewhere", size="sm") |
| q7 = gr.Button("Can you do better?", size="sm") |
| q8 = gr.Button("Ok I will take it", size="sm") |
|
|
| with gr.Row(): |
| msg_input = gr.Textbox( |
| placeholder="Type message in Hinglish or English...", |
| show_label=False, |
| scale=4, |
| container=False |
| ) |
| send_btn = gr.Button("Send", variant="primary", scale=1) |
|
|
| gr.HTML(""" |
| <div class="footer-note"> |
| Built by Nitesh Nankani · MuRIL + BERT + Claude Haiku + 9 Negotiation Books · HuggingFace · Gradio |
| </div> |
| """) |
|
|
| q1.click(fn=set_q1, outputs=msg_input) |
| q2.click(fn=set_q2, outputs=msg_input) |
| q3.click(fn=set_q3, outputs=msg_input) |
| q4.click(fn=set_q4, outputs=msg_input) |
| q5.click(fn=set_q5, outputs=msg_input) |
| q6.click(fn=set_q6, outputs=msg_input) |
| q7.click(fn=set_q7, outputs=msg_input) |
| q8.click(fn=set_q8, outputs=msg_input) |
|
|
| send_btn.click( |
| fn=chat, |
| 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], |
| outputs=[chatbot, intel_bar, current_offer, msg_input] |
| ).then(fn=update_price, inputs=[current_offer], outputs=[price_display]) |
|
|
| reset_btn.click( |
| fn=reset_chat, |
| inputs=[mrp], |
| outputs=[chatbot, intel_bar, current_offer] |
| ).then(fn=update_price, inputs=[current_offer], outputs=[price_display]) |
|
|
| if __name__ == "__main__": |
| demo.launch() |
|
|
