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


def build_doc_frame(df, idx=0):
    doc = df.iloc[0]
    # 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(doc_idx, query_str, K=5):
    # doc_idx is a choice option of (idx, text)
    idx = doc_idx[0] - 1
    newdoc = build_doc_frame(df, idx=idx)
    embeddings = get_doc_embeddings(newdoc)

    faiss.normalize_L2(embeddings)
    index = faiss.IndexFlatIP(768)
    index.add(embeddings)

    query_str = "Skade mellom kjøretøy"
    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}\t{sent}" for score, sent in pretty_results])
    top_k_str = f"Top {K} results for: {query_str}"
    underlines = "__" * 40

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


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

dropdown_opts = [(idx + 1, f"\t{doc.summary[0][:60]}...") for idx, doc in df.iterrows()]

iface = gr.Interface(
    fn=faiss_search,
    inputs=[
        gr.Dropdown(label="Select a court case", choices=dropdown_opts),
        gr.Textbox(lines=2, placeholder="Your query here..."),
        gr.Slider(minimum=1, maximum=10, label="Number of matches", value=5),
    ],
    outputs="text",
    title="Lovdata rettsavgjørelser - semantisk søk",
    description="Velg en rettsak og søk for å hente ut lignende setninger i saken",
)

iface.launch()