app.py

from sentence_transformers import SentenceTransformer, util
import torch
import os
import gradio as gr
import json
import re
import numpy as np
import pickle
from datetime import datetime
from huggingface_hub import Repository
from datasets import load_dataset

############# Read in the data #############

access_token_1 = os.environ.get("HF_token")

dataset = load_dataset("acd424/tribunal_data", use_auth_token=access_token_1)


embeddings = dataset["train"]["embed"]

corpus = dataset["train"]["reason_text"]

files = dataset["train"]["file_name"]

all_cats = dataset["train"]["all_cats"]

print("The data has loaded")

'''# for saving a log
DATASET_REPO_URL = "https://huggingface.co/datasets/acd424/tribunal_data"
DATA_FILENAME = "queries_and_responces.txt"
DATA_FILE = os.path.join("data", DATA_FILENAME)
repo = Repository(
    local_dir="data", clone_from=DATASET_REPO_URL, use_auth_token=access_token_1
)
'''
################### Functions ##########################


def semantic_search(query, corpus=corpus, corpus_embeddings=embeddings, k=5):
    """Performs search to find vector in corpus closest to query (using cosine similarity)
    and returns the top k resluts

    Parameters
    ----------
    query: str
        The query the user wants to search with
    corpus: list
        Each item in the list is the text from a chunk
    corpus_embedding: numpy.ndarray
        The word-embedding vector representation of the chunk
    k: int
        The numer of results to print (default is 5)

    Returns
    -------
    Prints output displaying top k results
    """

    top_k = min(k, len(corpus_embeddings))

    query_embedding = embedder.encode(query, convert_to_tensor=True)

    # We use cosine-similarity and torch.topk to find the highest 5 scores
    cos_scores = util.cos_sim(query_embedding, corpus_embeddings)[0]
    top_results = torch.topk(cos_scores, k=top_k)
    highest_cos_score = cos_scores[list(np.argsort(cos_scores))[::-1][0]]

    if highest_cos_score < 0.5:
        final_string = "The query is either not detailed enough or is perhaps not an appropriate query"
    else:
        final_string = ""
        n = 1
        for score, idx in zip(top_results[0], top_results[1]):
            final_string = final_string + f" Match: {n} \n"
            final_string = (
                final_string
                + f"https://www.gov.uk/employment-tribunal-decisions/{corpus[idx]}"
            )
            final_string = final_string + "\n   ======== With Catergories =========  \n"
            final_string = final_string + all_cats[idx]
            final_string = (
                final_string
                + f"\n   ============================================ {cos_scores[list(np.argsort(cos_scores))[::-1][n-1]]} \n\n\n"
            )
            n += 1

    return final_string


def produce_tribunal_out(query, corpus=files, tribunal_embeddings=embeddings):
    # get context
    context_string = semantic_search(
        query=query, corpus=corpus, corpus_embeddings=tribunal_embeddings
    )
    # get time
    time_now = datetime.now().strftime("%d/%m/%Y %H:%M:%S")
    print(f"{time_now}|{query}\n")
    '''# write to file

    with open(DATA_FILE, "a") as f:
        f.write(f"{time_now}|{query}\n")
    commit_url = repo.push_to_hub()
    print(commit_url)
    '''
    return gr.update(value=context_string)


############### Specify models

embedder = SentenceTransformer("all-MiniLM-L6-v2")

############### The Appp

with gr.Blocks() as demo:
    gr.Markdown("Employment tribunal demo")

    # text input from user
    inp = gr.Textbox(
        placeholder="Type your employment tribunal query here - describe your problem",
        label="Tribunal Query",
    )

    # initiate the functions
    process_btn = gr.Button("Search records from UK employment tribunals")

    Suggested_text = gr.Textbox(
        value="Suggestions will appear here", label="", lines=20
    )

    Suggested_text.style(show_copy_button=True)

    process_btn.click(
        fn=produce_tribunal_out,
        inputs=[inp],
        outputs=Suggested_text,
    )

demo.launch(auth=("admin", "dataisking"))


##########################################