chatbot.py

from transformers import T5Tokenizer, T5ForConditionalGeneration
from sentence_transformers import SentenceTransformer
from pinecone import Pinecone

device = 'cpu'

# Calling the pinecone api
pc = Pinecone(api_key='89eeb534-da10-4068-92f7-12eddeabe1e5')

# Connect to the Pinecone index for querying and storing vectors
index_name = 'abstractive-question-answering'
index = pc.Index(index_name)

# Load the retriever model for sentence embeddings and the T5 model for text generation
def load_models():
    print("Loading models...")   
    retriever = SentenceTransformer("flax-sentence-embeddings/all_datasets_v3_mpnet-base")
    tokenizer = T5Tokenizer.from_pretrained('t5-small')
    generator = T5ForConditionalGeneration.from_pretrained('t5-base').to(device)
    return retriever, generator, tokenizer
    print("Done loading models") 

retriever, generator, tokenizer = load_models()

def process_query(query):    
    print("Processing...") 
    # Encode the query into a vector for semantic search using SentenceTransformer
    xq = retriever.encode([query]).tolist()
    # Query the Pinecone index for the most similar vector to the query
    xc = index.query(vector=xq, top_k=1, include_metadata=True)
    
    print("Pinecone response:", xc)

    # Concatenates the original question with the context extracted from the matched metadata
    if 'matches' in xc and isinstance(xc['matches'], list):
        context = [m['metadata']['Output'] for m in xc['matches']]
        context_str = " ".join(context)
        formatted_query = f"answer the question: {query} context: {context_str}"

    # If the context is longer than 5 lines, return the context extracted from Pinecone directly
    output_text = context_str
    if len(output_text.splitlines()) > 5:
        return output_text

    # If none, then it will return that it was not covered in the student manual
    if output_text.lower() == "none":
        return "The topic is not covered in the student manual."

    # Tokenizes the formatted query
    inputs = tokenizer.encode(formatted_query, return_tensors="pt", max_length=512, truncation=True).to(device)
    # Generates an answer using the t5 model
    ids = generator.generate(inputs, num_beams=2, min_length=10, max_length=60, repetition_penalty=1.2)
    # Decodes the answer to make it readable for the user
    answer = tokenizer.decode(ids[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
    

    # If it has this words, it will just paste the output from the extracted meta-data output from pinecone
    nli_keywords = ['not_equivalent', 'not_entailment', 'entailment', 'neutral', 'not_enquiry']    
    if any(keyword in answer.lower() for keyword in nli_keywords):
        return context_str  

    # returns the answer
    return answer