llm.py

from transformers import pipeline
import torch
from langchain.prompts import ChatPromptTemplate
from langchain.schema.output_parser import StrOutputParser
from langchain_huggingface import HuggingFacePipeline

# Initialize HF pipeline for text generation
text_generator = pipeline(
    "text-generation",  # Task type
    model="google/gemma-3n-E2B-it",
    # model="google/gemma-3n-e4b-it",
    # model="Qwen/Qwen3-Embedding-0.6B",
    # device="cuda" if torch.cuda.is_available() else "cpu",
    device= "cpu",
    torch_dtype=torch.bfloat16,
    max_new_tokens=500  # Limit output length
)

# Wrap pipeline for LangChain compatibility
model = HuggingFacePipeline(pipeline=text_generator)

def generate_sentences(topic, n=1):
    prompt = ChatPromptTemplate.from_template(
        "You are a helpful assistant. Generate exactly {n} simple sentences about the topic: {topic}. "
        "Each sentence must be in English and appropriate for all audiences. "
        "Return each sentence on a new line without any numbering or bullets"
    )
    chain = prompt | model | StrOutputParser()
    response = chain.invoke({"topic": topic, "n": n})
    
    # Enhanced filtering
    return [
        line.strip() for line in response.splitlines() 
        if (line.strip() 
            and not line.startswith(("###", "Instruction", "Output Format"))
            and len(line.split()) <= 15  # Word limit enforcement
        )
    ][:n]