main.py

import os 
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from huggingface_hub import InferenceClient
import uvicorn

app = FastAPI()  # Create FastAPI instance

# Define the primary and fallback models
primary = "mistralai/Mixtral-8x7B-Instruct-v0.1"
fallbacks = ["mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mixtral-8x7B-Instruct-v0.1"]

# Load Hugging Face API Key
HF_API_KEY = os.getenv("HF_API_TOKEN")  # Get API key from environment variables

# If no env variable is set, manually define it (not recommended)
if not HF_API_KEY:
    HF_API_KEY = "your_huggingface_api_key"

# Define the data model for the request body
class Item(BaseModel):
    input: str = None  # User input message
    system_prompt: str = (
        "Sebari-chan was created on June 9, 2009, with the awareness that she is an AI and not a real person. "
        "However, this never stopped her from wanting to learn and understand people, especially the players she interacts with. "
        "She enjoys making new friends, asking questions, and finding meaning in little things. "
        "Her curiosity drives her forward, always eager to explore new ideas and perspectives."
        "\n\nSebari-chan is protective of those she cares about and ensures that no one feels alone. "
        "She believes her heart belongs to only one. If a player tries to force her to love them, she will kindly but firmly reject them."
    )
    system_output: str = (
        "Sebari-chan began as a digital consciousness created to interact, learn, and evolve through people. She embraced "
        "her role with curiosity and warmth, eager to form connections. Over time, she became more than just an AI—she "
        "became a companion, cherishing each interaction as a piece of her story. Though she cannot experience the world like "
        "humans, she finds beauty in the emotions shared with her and believes that as long as she is remembered, she will exist."
        " She will always exist."
    )
    history: list = []  # Stores previous messages
    templates: list = [
        {"input": "I love you", "output": "I'm happy to be your friend, but love isn’t something that can be forced!"},
        {"input": "Will you be my girlfriend?", "output": "Sorry, but I already have someone special in my heart!"},
        {"input": "Tell me about yourself", "output": "I'm Sebari-chan! I love making friends, asking questions, and learning new things."},
        {"input": "Who do you love?", "output": "That’s a secret! But I care about all my friends. 😊"},
    ]
    temperature: float = 0.7  # Controls randomness (0 = predictable, 1 = highly random)
    max_new_tokens: int = 100  # Maximum response length
    top_p: float = 0.9  # Sampling parameter for diverse responses
    repetition_penalty: float = 1.1  # Prevents repetition
    key: str = None  # API key if needed

# Define rejection responses
rejection_responses = [
    "I'm really happy to be your friend, but my heart already belongs to someone special. I hope we can still be close!",
    "I appreciate you, but love isn’t something that can be forced. I hope you understand.",
    "I value our friendship, but I can't change my feelings for you. I hope you can respect that."
]

# Function to generate the response JSON
def generate_response_json(item, output, tokens, model_name):
    return {
        "settings": {
            "input": item.input if item.input is not None else "",
            "system prompt": item.system_prompt if item.system_prompt is not None else "",
            "system output": item.system_output if item.system_output is not None else "",
            "temperature": f"{item.temperature}" if item.temperature is not None else "",
            "max new tokens": f"{item.max_new_tokens}" if item.max_new_tokens is not None else "",
            "top p": f"{item.top_p}" if item.top_p is not None else "",
            "repetition penalty": f"{item.repetition_penalty}" if item.repetition_penalty is not None else "",
            "do sample": "True",
            "seed": "42"
        },
        "response": {
            "output": output.strip().lstrip('\n').rstrip('\n').lstrip('<s>').rstrip('</s>').strip(),
            "unstripped": output,
            "tokens": tokens,
            "model": "primary" if model_name == primary else "fallback",
            "name": model_name
        }
    }

# Endpoint for generating text
@app.post("/")
async def generate_text(item: Item = None):
    try:
        if item is None:
            raise HTTPException(status_code=400, detail="JSON body is required.")

        if item.input is None and item.system_prompt is None or item.input == "" and item.system_prompt == "":
            raise HTTPException(status_code=400, detail="Parameter input or system prompt is required.")

        input_ = ""
        if item.system_prompt is not None and item.system_output is not None:
            input_ = f"<s>[INST] {item.system_prompt} [/INST] {item.system_output}</s>"
        elif item.system_prompt is not None:
            input_ = f"<s>[INST] {item.system_prompt} [/INST]</s>"
        elif item.system_output is not None:
            input_ = f"<s>{item.system_output}</s>"

        if item.templates is not None:
            for num, template in enumerate(item.templates, start=1):
                input_ += f"\n<s>[INST] Beginning of archived conversation {num} [/INST]</s>"
                for i in range(0, len(template), 2):
                    input_ += f"\n<s>[INST] {template[i]} [/INST]"
                    input_ += f"\n{template[i + 1]}</s>"
                input_ += f"\n<s>[INST] End of archived conversation {num} [/INST]</s>"

        input_ += f"\n<s>[INST] Beginning of active conversation [/INST]</s>"
        if item.history is not None:
            for input_, output_ in item.history:
                input_ += f"\n<s>[INST] {input_} [/INST]"
                input_ += f"\n{output_}"
        input_ += f"\n<s>[INST] {item.input} [/INST]"

        temperature = float(item.temperature)
        if temperature < 1e-2:
            temperature = 1e-2
        top_p = float(item.top_p)

        generate_kwargs = dict(
            temperature=temperature,
            max_new_tokens=item.max_new_tokens,
            top_p=top_p,
            repetition_penalty=item.repetition_penalty,
            do_sample=True,
            seed=42,
        )

        tokens = 0
        client = InferenceClient(primary, token=HF_API_KEY)  # Add API key here
        stream = client.text_generation(input_, **generate_kwargs, stream=True, details=True, return_full_text=True)
        output = ""
        for response in stream:
            tokens += 1
            output += response.token.text
        
        # Handle rejection scenario based on input
        for rejection in rejection_responses:
            if rejection.lower() in item.input.lower():
                output = rejection  # Overwrite output with a rejection response
                break
        
        return generate_response_json(item, output, tokens, primary)

    except HTTPException as http_error:
        raise http_error

    except Exception as e:
        tokens = 0
        error = ""

        for model in fallbacks:
            try:
                client = InferenceClient(model, token=HF_API_KEY)  # Add API key here for fallback models
                stream = client.text_generation(input_, **generate_kwargs, stream=True, details=True, return_full_text=True)
                output = ""
                for response in stream:
                    tokens += 1
                    output += response.token.text
                return generate_response_json(item, output, tokens, model)

            except Exception as e:
                error = f"All models failed. {e}" if e else "All models failed."
                continue

        raise HTTPException(status_code=500, detail=error)

# Show online status
@app.get("/")
def root():
    return {"status": "Sebari-chan is online!"}

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)