app.py

import streamlit as st
import pandas as pd

from backend import inference
from backend.config import MODELS_ID, QA_MODELS_ID, SEARCH_MODELS_ID

st.title('Demo using Flax-Sentence-Tranformers')

st.sidebar.title('Tasks')
menu = st.sidebar.radio("", options=["Sentence Similarity", "Asymmetric QA", "Search / Cluster"], index=0)

st.markdown('''

Hi! This is the demo for the [flax sentence embeddings](https://huggingface.co/flax-sentence-embeddings) created for the **Flax/JAX community week 🤗**. We are going to use three flax-sentence-embeddings models: a **distilroberta base**, a **mpnet base** and a **minilm-l6**. All were trained on all the dataset of the 1B+ train corpus with the v3 setup.

''')

if menu == "Sentence Similarity":
    st.header('Sentence Similarity')
    st.markdown('''
**Instructions**: You can compare the similarity of a main text with other texts of your choice. In the background, we'll create an embedding for each text, and then we'll use the cosine similarity function to calculate a similarity metric between our main sentence and the others.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')
    select_models = st.multiselect("Choose models", options=list(MODELS_ID), default=list(MODELS_ID)[0])

    anchor = st.text_input(
        'Please enter here the main text you want to compare:'
    )

    n_texts = st.number_input(
        f'''How many texts you want to compare with: '{anchor}'?''',
        value=2,
        min_value=2)

    inputs = []

    for i in range(int(n_texts)):
        input = st.text_input(f'Text {i + 1}:')

        inputs.append(input)

    if st.button('Tell me the similarity.'):
        results = {model: inference.text_similarity(anchor, inputs, model, MODELS_ID) for model in select_models}
        df_results = {model: results[model] for model in results}

        index = [f"{idx + 1}:{input[:min(15, len(input))]}..." for idx, input in enumerate(inputs)]
        df_total = pd.DataFrame(index=index)
        for key, value in df_results.items():
            df_total[key] = list(value['score'].values)

        st.write('Here are the results for selected models:')
        st.write(df_total)
        st.write('Visualize the results of each model:')
        st.line_chart(df_total)
elif menu == "Asymmetric QA":
    st.header('Asymmetric QA')
    st.markdown('''
**Instructions**: You can compare the Answer likeliness of a given Query with answer candidates of your choice. In the background, we'll create an embedding for each answers, and then we'll use the cosine similarity function to calculate a similarity metric between our query sentence and the others.
`mpnet_asymmetric_qa` model works best for hard negative answers or distinguishing similar queries due to separate models applied for encoding questions and answers.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')

    select_models = st.multiselect("Choose models", options=list(QA_MODELS_ID), default=list(QA_MODELS_ID)[0])

    anchor = st.text_input(
        'Please enter here the query you want to compare with given answers:',
        value="What is the weather in Paris?"
    )

    n_texts = st.number_input(
        f'''How many answers you want to compare with: '{anchor}'?''',
        value=10,
        min_value=2)

    inputs = []

    defaults = ["It is raining in Paris right now with 70 F temperature.", "What is the weather in Berlin?", "I have 3 brothers."]
    for i in range(int(n_texts)):
        input = st.text_input(f'Answer {i + 1}:', value=defaults[i] if i < len(defaults) else "")

        inputs.append(input)

    if st.button('Tell me Answer likeliness.'):
        results = {model: inference.text_similarity(anchor, inputs, model, QA_MODELS_ID) for model in select_models}
        df_results = {model: results[model] for model in results}

        index = [f"{idx + 1}:{input[:min(15, len(input))]}..." for idx, input in enumerate(inputs)]
        df_total = pd.DataFrame(index=index)
        for key, value in df_results.items():
            df_total[key] = list(value['score'].values)

        st.write('Here are the results for selected models:')
        st.write(df_total)
        st.write('Visualize the results of each model:')
        st.line_chart(df_total)

elif menu == "Search / Cluster":
    st.header('Search / Cluster')
    st.markdown('''
**Instructions**: Make a query for anything related to "Python" and the model you choose will return you similar queries.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')

    select_models = st.multiselect("Choose models", options=list(SEARCH_MODELS_ID), default=list(SEARCH_MODELS_ID)[0])

    anchor = st.text_input(
        'Please enter here your query about "Python", we will look for similar ones:',
        value="How do I sort a dataframe by column"
    )

    n_texts = st.number_input(
        f'''How many similar queries you want?''',
        value=3,
        min_value=2)

    if st.button('Give me my search.'):
        results = {model: inference.text_search(anchor, n_texts, model, QA_MODELS_ID) for model in select_models}
        st.table(pd.DataFrame(results[select_models[0]]).T)

    if st.button('3D Clustering of search result (new window)'):
        fig = inference.text_cluster(anchor, 1000, select_models[0], QA_MODELS_ID)
        fig.show()