app.py

import os
import gradio as gr
from sentence_transformers import SentenceTransformer, CrossEncoder, util
from transformers import pipeline
import torch
import pickle
import pandas as pd
import gradio as gr


##Speech Recognition
asr = pipeline("automatic-speech-recognition", "facebook/wav2vec2-base-960h")
def speech_to_text(speech):
    text = asr(speech)["text"]
    return text

bi_encoder = SentenceTransformer("multi-qa-MiniLM-L6-cos-v1")
cross_encoder = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")
corpus=pd.read_pickle("corpus.pkl")
corpus_embeddings=pd.read_pickle("corpus_embeddings_cpu.pkl")


def search(query,top_k=100):
    print("Top 3 Answer by the NSE:")
    print()
    ans=[]
    ##### Sematic Search #####
    # Encode the query using the bi-encoder and find potentially relevant passages
    question_embedding = bi_encoder.encode(query, convert_to_tensor=True)
    hits = util.semantic_search(question_embedding, corpus_embeddings, top_k=top_k)
    hits = hits[0]  # Get the hits for the first query
    ##### Re-Ranking #####
    # Now, score all retrieved passages with the cross_encoder
    cross_inp = [[query, corpus[hit['corpus_id']]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp)
    # Sort results by the cross-encoder scores
    for idx in range(len(cross_scores)):
        hits[idx]['cross-score'] = cross_scores[idx]
    hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
    
    for idx, hit in enumerate(hits[0:3]):
        ans.append(corpus[hit['corpus_id']])
    return ans[0],ans[1],ans[2]


demo = gr.Blocks()
with demo:
    audio_file = gr.inputs.Audio(source="microphone", type="filepath")
    b1 = gr.Button("Recognize Speech")
    text = gr.Textbox()
    b1.click(speech_to_text, inputs=audio_file, outputs=text)
    b2 = gr.Button("Ask Wiki")
    print(text)
    out1 = gr.Textbox()
    out2 = gr.Textbox()
    out3 = gr.Textbox()
    b2.click(search, inputs=text, outputs=[out1,out2,out3])
demo.launch(debug=True)