app.py

import gradio as gr

from sentence_transformers import SentenceTransformer
import duckdb
from huggingface_hub import get_token

from sentence_transformers import SentenceTransformer
from sentence_transformers.models import StaticEmbedding
import duckdb

# Initialize a StaticEmbedding module
static_embedding = StaticEmbedding.from_model2vec("minishlab/potion-base-8M")
model = SentenceTransformer(modules=[static_embedding])

dataset_name = "smol-blueprint/fineweb-bbc-news-text-embeddings"
embedding_column = "embedding"

duckdb.sql(
    query=f"""
    INSTALL vss;
    LOAD vss;
    CREATE TABLE embeddings AS 
    SELECT *, {embedding_column}::float[{model.get_sentence_embedding_dimension()}] as embedding_float 
    FROM 'hf://datasets/{dataset_name}/**/*.parquet';
    CREATE INDEX my_hnsw_index ON embeddings USING HNSW (embedding_float) WITH (metric = 'cosine');
"""
)

def similarity_search(query: str, k: int = 5):
    embedding = model.encode(query).tolist()
    return duckdb.sql(
        query=f"""
        SELECT url, chunk, array_cosine_distance(embedding_float, {embedding}::FLOAT[{model.get_sentence_embedding_dimension()}]) as distance 
        FROM embeddings 
        ORDER BY distance 
        LIMIT {k};
    """
    ).to_df()

with gr.Blocks() as demo:
    gr.Markdown("""# Vector Search Hub Datasets
                
                Part of [smol blueprint](https://github.com/huggingface/smol-blueprint) - a smol blueprint for AI development, focusing on applied examples of RAG, information extraction, analysis and fine-tuning in the age of LLMs. """)
    query = gr.Textbox(label="Query")
    k = gr.Slider(1, 50, value=5, label="Number of results")
    btn = gr.Button("Search")
    results = gr.Dataframe(headers=["url", "chunk", "distance"])
    btn.click(fn=similarity_search, inputs=[query, k], outputs=[results])
    

demo.launch()