main.py

from fastapi import FastAPI, HTTPException
from fastapi.responses import StreamingResponse
from fastapi.responses import JSONResponse
from pydantic import BaseModel
from huggingface_hub import InferenceClient
import uvicorn
from typing import Generator
import json  # Asegúrate de que esta línea esté al principio del archivo
import nltk
import os
from transformers import pipeline


nltk.data.path.append(os.getenv('NLTK_DATA'))

app = FastAPI()

# Initialize the InferenceClient with your model
client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.2")

# summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6")

summarizer = pipeline("summarization", model="sshleifer/distilbart-cnn-12-6")

class Item(BaseModel):
    prompt: str
    history: list
    system_prompt: str
    temperature: float = 0.8
    max_new_tokens: int = 12000
    top_p: float = 0.15
    repetition_penalty: float = 1.0

def format_prompt(current_prompt, history):
    formatted_history = "<s>"
    for entry in history:
        if entry["role"] == "user":
            formatted_history += f"[USER] {entry['content']} [/USER]"
        elif entry["role"] == "assistant":
            formatted_history += f"[ASSISTANT] {entry['content']} [/ASSISTANT]"
    formatted_history += f"[USER] {current_prompt} [/USER]</s>"
    return formatted_history


def generate_stream(item: Item) -> Generator[bytes, None, None]:
    formatted_prompt = format_prompt(f"{item.system_prompt}, {item.prompt}", item.history)
    # Estimate token count for the formatted_prompt
    input_token_count = len(nltk.word_tokenize(formatted_prompt))  # NLTK tokenization

    # Ensure total token count doesn't exceed the maximum limit
    max_tokens_allowed = 32768
    max_new_tokens_adjusted = max(1, min(item.max_new_tokens, max_tokens_allowed - input_token_count))

    generate_kwargs = {
        "temperature": item.temperature,
        "max_new_tokens": max_new_tokens_adjusted,
        "top_p": item.top_p,
        "repetition_penalty": item.repetition_penalty,
        "do_sample": True,
        "seed": 42,
    }

    # Stream the response from the InferenceClient
    for response in client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True):
        # This assumes 'details=True' gives you a structure where you can access the text like this
        chunk = {
            "text": response.token.text,
            "complete": response.generated_text is not None  # Adjust based on how you detect completion
        }
        yield json.dumps(chunk).encode("utf-8") + b"\n"


class SummarizeRequest(BaseModel):
    text: str

@app.post("/generate/")
async def generate_text(item: Item):
    # Stream response back to the client
    return StreamingResponse(generate_stream(item), media_type="application/x-ndjson")

def split_text(text, max_size=1000):
    # Splits the text into chunks of approximately `max_size` words
    words = text.split()
    for i in range(0, len(words), max_size):
        yield ' '.join(words[i:i+max_size])

def summarize_large_text(text):
    chunks = list(split_text(text))
    summaries = [summarizer(chunk, max_length=500, min_length=100, do_sample=False) for chunk in chunks]
    combined_summary = ' '.join(sum[0]['summary_text'] for sum in summaries)
    return combined_summary

@app.post("/summarize")
async def summarize_text(request: SummarizeRequest):
    try:
        # Adjusting summarization for very large texts
        summarized_text = summarize_large_text(request.text)
        return JSONResponse(content={"summary": summarized_text})
    except Exception as e:
        # Handle exceptions that could arise during summarization
        raise HTTPException(status_code=500, detail=str(e))

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