app.py

from sentence_transformers import SentenceTransformer, CrossEncoder, util
import torch
import pickle
import pandas as pd


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


def search(query, top_k=100):
    print("Top 5 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[0:5]):
        ans.append(corpus[hit['corpus_id']])
    return ans[0], ans[1], ans[2], ans[3], ans[4]


exp = ["Who is steve jobs?", "What is coldplay?", "What is a turing test?",
       "What is the most interesting thing about our universe?", "What are the most beautiful places on earth?"]

desc = "This is a semantic search engine powered by SentenceTransformers (Nils_Reimers) with a retrieval and reranking system on Wikipedia corous. This will return the top 5 results. So Quest on with Transformers."

inp = gr.inputs.Textbox(lines=1, placeholder=None, default="", label="search you query here")
out = gr.outputs.Textbox(type="auto", label="search results")

iface = gr.Interface(fn=search, inputs=inp, outputs=[out, out, out, out, out], examples=exp, article=desc,
                     title="Search Engine", theme="huggingface", layout='vertical')
iface.launch()