main.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import pipeline, AutoTokenizer,T5ForConditionalGeneration


tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-base")
model = T5ForConditionalGeneration.from_pretrained("google-t5/t5-base")


app = FastAPI()
summarizer = pipeline("summarization", model="facebook/bart-large-cnn")


class TextInput(BaseModel):
    text: str


@app.get("/")
async def root():
    return {"message": "Welcome to the Text Summarization API!"}

@app.post("/summarize")
async def summarize_text(input: TextInput):
    # Count approximate number of words (could be improved with tokenizer count)
    word_count = len(input.text.split())
    
    # Set dynamic parameters based on input length
    if word_count < 50:
        max_length = max(10, word_count // 2)  # Half the original length, minimum 10
        min_length = max(3, word_count // 4)   # Quarter the original length, minimum 3
    elif word_count < 200:
        max_length = max(50, word_count // 3)
        min_length = max(15, word_count // 6)
    else:
        max_length = max(100, word_count // 4)
        min_length = max(30, word_count // 8)
    
    # Prevent max_length from being too large (BART has token limits)
    max_length = min(max_length, 1024)
    
    # Generate summary with dynamic parameters
    summary = summarizer(
        input.text, 
        max_length=max_length, 
        min_length=min_length, 
        do_sample=True,
        temperature=0.7,
        num_beams=4
    )
    
    return {
        "summary": summary[0]["summary_text"]
    }
    

@app.post("/translateFrench")
async def translate(input: TextInput):
    # Step 1: Prefix the task for the model
    prefixed_text = "translate English to French: " + input.text

    # Step 2: Tokenize the input
    inputs = tokenizer(prefixed_text, return_tensors="pt", truncation=True)

    # Step 3: Adjust generation parameters
    input_length = inputs.input_ids.shape[1]
    max_length = min(512, input_length * 2)  # 2x input length but not more than 512
    min_length = int(input_length * 1.1)     # at least 10% longer than input

    # Step 4: Generate translation
    outputs = model.generate(
        **inputs,
        max_length=max_length,
        min_length=min_length,
        num_beams=5,
        length_penalty=1.2,
        early_stopping=True,
        no_repeat_ngram_size=2
    )

    # Step 5: Decode result
    translated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

    return {"translated_text": translated_text}