app.py

import faiss
import gradio as gr
import numpy as np
import pandas as pd
from datasets import load_dataset
from sentence_transformers import SentenceTransformer

idx = 0
index = None
newdoc = None

dataset = load_dataset("tollefj/rettsavgjoerelser_100samples_embeddings")
model = SentenceTransformer("NbAiLab/nb-sbert-base")
df = dataset["train"].to_pandas()


def build_doc_frame(df, idx):
    doc = df.iloc[idx]
    # as df:
    doc_df = pd.DataFrame(doc).T
    # keep only sentences + embedding:
    doc_df = doc_df[["url", "sentences", "embedding"]]
    # unpack the sentences and embedding in separate rows
    doc_df = doc_df.explode(["sentences", "embedding"])
    return doc_df


def get_doc_embeddings(doc):
    return np.array(doc.embedding.tolist(), dtype="float32")


def faiss_search(query_str, K=5):
    global idx
    global index
    global newdoc
    # find idx from url:
    # doc_idx = df[df.url == doc_url].index[0]
    # idx = int(doc_idx)
    target_emb = model.encode([query_str])
    target_emb = np.array([target_emb.reshape(-1)])
    faiss.normalize_L2(target_emb)

    D, I = index.search(np.array(target_emb), K)
    print(list(zip(D[0], I[0])))

    # prettyprint the results:
    pretty_results = []
    for idx, score in zip(I[0], D[0]):
        pretty_results.append((round(float(score), 3), newdoc.iloc[idx].sentences))
    pretty_results_str = "\n".join([f"Score: {score}\t\t{sent}" for score, sent in pretty_results])
    top_k_str = f"Top {K} results for: {query_str}"

    # return str:
    return f"{top_k_str}\n{pretty_results_str}"


# def DropdownSummary():
#     next_opts = df.iloc[idx].summary.tolist()
#     return gr.Dropdown.update(choices=next_opts, label="Velg fra oppsummeringene")


dropdown_opts = [doc.url for idx, doc in df.iterrows()]

with gr.Blocks() as demo:
    gr.HTML(
        """
        <h1>Lovdata rettsavgjørelser - semantisk søk</h1>
        """
    )

    def on_selection_change(selected_case):
        global idx
        global index
        global newdoc
        idx = df[df.url == selected_case].index[0]
        print("Selection changed!")
        print(selected_case)

        newdoc = build_doc_frame(df, idx)
        embeddings = get_doc_embeddings(newdoc)
        faiss.normalize_L2(embeddings)
        index = faiss.IndexFlatIP(768)
        index.add(embeddings)

        summary = df.iloc[idx].summary.tolist()
        # make a nice html-formatted ul-li list:
        summary_html = "<ul>" + "".join([f"<li>{sent}</li>" for sent in summary]) + "</ul>"
        # summary_dropdown.update(choices=summary, label="Velg fra oppsummeringene")
        url_html = f"<a href='{selected_case}' target='_blank'>{selected_case}</a>"
        return summary_html, url_html

    with gr.Row():
        with gr.Column():
            case_dropdown = gr.Dropdown(label="Velg en rettsavgjørelse", choices=dropdown_opts)
            summary_html = gr.HTML(label="Predefinert oppsummering", placeholder="<p>Velg en sak først<p>")
            case_url = gr.HTML(label="URL til rettsavgjørelse", placeholder="https://lovdata.no/...")
        with gr.Column():
            query = gr.Textbox(
                label="Søk etter setninger",
                lines=1,
                placeholder="Kollisjon mellom to kjøretøy.",
            )
            k_slider = gr.Slider(minimum=1, maximum=10, label="Antall treff", value=5, step=1)
            search_btn = gr.Button("Søk")

    output = gr.Textbox(label="Resultater", lines=10)

    case_dropdown.change(
        on_selection_change,
        inputs=[case_dropdown],
        outputs=[summary_html, case_url],
    )

    search_btn.click(faiss_search, inputs=[query, k_slider], outputs=[output])

    # clear_btn.click(None, inputs=[None, None], outputs=None)
    # search_btn.click(faiss_search, inputs=[None, None, None], outputs=["text"])
    # search_btn.click(faiss_search, inputs=[idx, query, k_slider], outputs=["text"])

demo.launch()