Initial Space setup

app.py

@@ -2,37 +2,71 @@ import gradio as gr
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
 from peft import PeftModel
 
-# 1) Load the original base model & tokenizer
+# 1) Base model & tokenizer
 BASE_MODEL = "facebook/blenderbot-400M-distill"
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
 base_model = AutoModelForSeq2SeqLM.from_pretrained(BASE_MODEL)
 
-# 2) Load your fine-tuned LoRA adapter on top
+# 2) Attach your LoRA adapter
 ADAPTER_REPO = "abinashnp/bayedger-chatbot"
 model = PeftModel.from_pretrained(base_model, ADAPTER_REPO)
 
-# 3) Wrap that in a text2text pipeline
+# 3) Build the text2text pipeline (no explicit device arg)
 chatbot = pipeline(
     "text2text-generation",
     model=model,
     tokenizer=tokenizer,
-    device_map="auto",       # leave out device arg when using accelerate device_map
+    # device_map="auto"  # only if you use Accelerate; otherwise remove
+)
+
+# 4) System prompt (context) that always precedes user questions
+SYSTEM_PROMPT = (
+    "You are BayEdger’s AI assistant. You only answer FAQs about BayEdger’s "
+    "services, pricing, and contact info. If you don’t know the answer, "
+    "you must say exactly:\n"
+    '"Sorry, I don’t have that info—please contact [email protected]."\n\n'
+    "Here is what you should know about BayEdger:\n"
+    "- AI‐powered websites and automation\n"
+    "- Chatbots, email agents, process automation, analytics, content gen\n"
+    "- Clear pricing tiers: Basic site ($400), Chatbot ($750+50/mo), Email ($1k+100/mo), etc.\n"
+    "- Starter/Growth/Premium bundles\n"
+    "- Contact: [email protected], +1‐234‐559‐87994, 13 Madison St, NY\n\n"
 )
 
 def respond(query):
+    # 5) Compose full prompt
+    prompt = (
+        SYSTEM_PROMPT
+        f"question: {query}\n"
+        "answer:"
+    )
+
+    # 6) Generate
     out = chatbot(
-        f"question: {query} answer:",
-        max_new_tokens=150,
-        temperature=1.0,
-        top_p=0.9,
-        repetition_penalty=1.1,
-        num_beams=1
+        prompt,
+        max_new_tokens=128,
+        do_sample=False,
+        num_beams=2,
+        early_stopping=True,
+        pad_token_id=tokenizer.eos_token_id
     )[0]["generated_text"]
-    return out
 
+    # 7) Strip off everything up through our "answer:" token
+    if "answer:" in out:
+        reply = out.split("answer:", 1)[1].strip()
+    else:
+        reply = out.strip()
+
+    # 8) Fallback: if the model didn’t produce anything substantial
+    if len(reply) < 15 or "don't know" in reply.lower() or "sorry" in reply.lower():
+        return "Sorry, I don’t have that info—please contact [email protected]."
+
+    return reply
+
+# 9) Gradio UI
 with gr.Blocks() as demo:
-    gr.Markdown("# 🤖 Bayedger FAQ Chatbot")
-    txt = gr.Textbox(label="Ask me anything")
+    gr.Markdown("# 🤖 BayEdger FAQ Chatbot")
+    txt = gr.Textbox(placeholder="Ask me about BayEdger…", label="Your question")
     out = gr.Textbox(label="Answer")
     txt.submit(respond, txt, out)