Removed requirements ... more pt 3

app.py

@@ -0,0 +1,131 @@
+from fastapi import FastAPI, HTTPException
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+from typing import List, Optional
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import re
+import os
+
+app = FastAPI()
+
+# Enable CORS for frontend access
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],  # Or specify your frontend domain
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# Hugging Face model config
+REPO_NAME = "jaydatech/phi3-finetuned-project"
+HF_TOKEN = os.getenv("HF_TOKEN")  # Load from Render environment variable
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+tokenizer = AutoTokenizer.from_pretrained(REPO_NAME, token=HF_TOKEN)
+model = AutoModelForCausalLM.from_pretrained(
+    REPO_NAME,
+    token=HF_TOKEN,
+    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+    device_map="auto"
+)
+
+# Message and request models
+class ChatMessage(BaseModel):
+    role: str
+    text: str
+
+class ChatRequest(BaseModel):
+    message: str
+    history: Optional[List[ChatMessage]] = []
+
+def is_farewell(message: str) -> bool:
+    farewells = ["bye", "goodbye", "see you", "farewell", "exit", "quit", "end"]
+    message_lower = message.lower().strip()
+    return any(farewell in message_lower for farewell in farewells)
+
+@app.post("/chat")
+async def chat(request: ChatRequest):
+    try:
+        history = request.history
+        user_message = request.message
+
+        if is_farewell(user_message):
+            return {
+                "response": "Goodbye! Feel free to chat again if you have more questions.",
+                "terminate": True
+            }
+
+        conversation = (
+            "<|system|>\nYou are an AI assistant for the Federal Reserve Bank of St. Louis. "
+            "Answer questions based ONLY on your knowledge of the Federal Reserve Bank of St. Louis. "
+            "If the answer is NOT in the training data, respond with: 'I don't think this information is available. Maybe rephrase for me!'. "
+            "Answer ONLY what the user asks. Do not volunteer information unless specifically requested. "
+            "Do NOT ask: 'How can I assist you today?' after every response you give. "
+            "Provide concise answers to the exact question asked and nothing more.\n"
+        )
+
+        seen_messages = set()
+        for msg in history:
+            if msg.role == "user" and msg.text.strip() not in seen_messages:
+                conversation += f"<|user|>\n{msg.text.strip()}\n"
+                seen_messages.add(msg.text.strip())
+            elif msg.role == "model":
+                conversation += f"<|assistant|>\n{msg.text.strip()}\n"
+
+        conversation += f"<|user|>\n{user_message.strip()}\n<|assistant|>"
+
+        inputs = tokenizer(conversation, return_tensors="pt", padding=True, truncation=True, max_length=4096).to(device)
+
+        with torch.no_grad():
+            outputs = model.generate(
+                **inputs,
+                max_new_tokens=130,
+                do_sample=True,
+                temperature=0.1,
+                top_k=5,
+                pad_token_id=tokenizer.eos_token_id
+            )
+
+        full_response = tokenizer.decode(outputs[0], skip_special_tokens=False)
+        assistant_response = ""
+
+        if "<|assistant|>" in full_response:
+            assistant_sections = full_response.split("<|assistant|>")
+            for section in reversed(assistant_sections):
+                cleaned = section.strip()
+                if cleaned:
+                    cleaned = re.split(
+                        r"(<\|user\|>|<\|system\|>|<\|assistant\|>|\nuser[:\s]|<\|endoftext\|>)",
+                        cleaned
+                    )[0]
+                    cleaned = re.sub(r"\n?(User|Assistant)\s*[:：\-–]\s*.*", "", cleaned, flags=re.IGNORECASE).strip()
+                    assistant_response = cleaned
+                    break
+
+        if not assistant_response:
+            assistant_response = "⚠️ Sorry, I couldn't generate a response."
+
+        assistant_response = re.sub(r'\*\* Instruction \*\*:.*?(?=\n\n|\n$|$)', '', assistant_response, flags=re.DOTALL)
+        assistant_response = re.sub(r'\*\* Instruction \*\*.*?(?=\n\n|\n$|$)', '', assistant_response, flags=re.DOTALL)
+        assistant_response = re.sub(r'\n{3,}', '\n\n', assistant_response).strip()
+        assistant_response = re.sub(r"(How can I assist you today\?|What else can I help you with\?|How can I help you today\?)", "", assistant_response, flags=re.IGNORECASE).strip()
+
+        def remove_repeated_sentences(response):
+            sentences = response.split(". ")
+            seen = set()
+            cleaned = []
+            for sentence in sentences:
+                if sentence not in seen:
+                    cleaned.append(sentence)
+                    seen.add(sentence)
+            return ". ".join(cleaned)
+
+        assistant_response = remove_repeated_sentences(assistant_response)
+
+        return {"response": assistant_response}
+
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))