app.py

import os
from functools import lru_cache
from typing import Optional

import gradio as gr
from dotenv import load_dotenv
from qdrant_client import QdrantClient, models
from sentence_transformers import SentenceTransformer
from huggingface_hub import list_models

load_dotenv()

URL = os.getenv("QDRANT_URL")
QDRANT_API_KEY = os.getenv("QDRANT_API_KEY")
sentence_embedding_model = SentenceTransformer("BAAI/bge-large-en")

print(URL)
print(QDRANT_API_KEY)
collection_name = "dataset_cards"
client = QdrantClient(
    url=URL,
    api_key=QDRANT_API_KEY,
)


# def convert_bytes_to_human_readable_size(bytes_size):
#     if bytes_size < 1024**2:
#         return f"{bytes_size / 1024:.2f} MB"
#     elif bytes_size < 1024**3:
#         return f"{bytes_size / (1024 ** 2):.2f} GB"
#     else:
#         return f"{bytes_size / (1024 ** 3):.2f} TB"


def format_time_nicely(time_str):
    return time_str.split("T")[0]


def format_results(results, show_associated_models=True):
    markdown = (
        "<h1 style='text-align: center;'>  &#x2728; Dataset Search Results  &#x2728;"
        " </h1> \n\n"
    )
    for result in results:
        hub_id = result.payload["id"]
        download_number = result.payload["downloads"]
        lastModified = result.payload["lastModified"]
        url = f"https://huggingface.co/datasets/{hub_id}"
        header = f"## [{hub_id}]({url})"
        markdown += header + "\n"

        markdown += f"**30 Day Download:** {download_number}"
        if lastModified:
            markdown += f" | **Last Modified:** {format_time_nicely(lastModified)} \n\n"
        else:
            markdown += "\n\n"
        markdown += f"{result.payload['section_text']} \n"
        if show_associated_models:
            if linked_models := get_models_for_dataset(hub_id):
                linked_models = [
                    f"[{model}](https://huggingface.co/{model})"
                    for model in linked_models
                ]
                markdown += (
                    "<details><summary>Models trained on this dataset</summary>\n\n"
                )
                markdown += "- " + "\n- ".join(linked_models) + "\n\n"
                markdown += "</details>\n\n"

    return markdown


@lru_cache(maxsize=100_000)
def get_models_for_dataset(id):
    results = list(iter(list_models(filter=f"dataset:{id}")))
    if results:
        results = list({result.id for result in results})
    return results


@lru_cache(maxsize=200_000)
def search(query: str, limit: Optional[int] = 10, show_linked_models: bool = False):
    query_ = sentence_embedding_model.encode(
        f"Represent this sentence for searching relevant passages:{query}"
    )
    results = client.search(
        collection_name="dataset_cards",
        query_vector=query_,
        limit=limit,
    )
    return format_results(results, show_associated_models=show_linked_models)


@lru_cache(maxsize=100_000)
def hub_id_qdrant_id(hub_id):
    matches = client.scroll(
        collection_name="dataset_cards",
        scroll_filter=models.Filter(
            must=[
                models.FieldCondition(key="id", match=models.MatchValue(value=hub_id)),
            ]
        ),
        limit=1,
        with_payload=True,
        with_vectors=False,
    )
    try:
        return matches[0][0].id
    except IndexError as e:
        raise gr.Error(
            f"Hub id {hub_id} not in the database. This could be because it is very new"
            " or because it doesn't have much documentation."
        ) from e


@lru_cache()
def recommend(hub_id, limit: Optional[int] = 10, show_linked_models=False):
    positive_id = hub_id_qdrant_id(hub_id)
    results = client.recommend(
        collection_name=collection_name, positive=[positive_id], limit=limit
    )
    return format_results(results, show_associated_models=show_linked_models)


def query(
    search_term,
    search_type,
    limit: Optional[int] = 10,
    show_linked_models: bool = False,
):
    if search_type == "Recommend similar datasets":
        return recommend(search_term, limit, show_linked_models)
    else:
        return search(search_term, limit, show_linked_models)


with gr.Blocks() as demo:
    gr.Markdown("## &#129303; Semantic Dataset Search")
    with gr.Row():
        gr.Markdown(
            "This Gradio app allows you to search for datasets based on their"
            " descriptions. You can either search for similar datasets to a given"
            " dataset or search for datasets based on a query. This is an early proof of concept. Feedback very welcome!"
        )
    with gr.Row():
        search_term = gr.Textbox(
            value="movie review sentiment",
            label="hub id i.e. IMDB or query i.e. movie review sentiment",
        )

    with gr.Row():
        with gr.Row():
            find_similar_btn = gr.Button("Search")
            search_type = gr.Radio(
                ["Recommend similar datasets", "Semantic Search"],
                label="Search type",
                value="Semantic Search",
                interactive=True,
            )
        with gr.Column():
            max_results = gr.Slider(
                minimum=1,
                maximum=50,
                step=1,
                value=10,
                label="Maximum number of results",
            )
            show_linked_models = gr.Checkbox(
                label="Show associated models",
                default=False,
            )

    results = gr.Markdown()
    find_similar_btn.click(
        query, [search_term, search_type, max_results, show_linked_models], results
    )


demo.launch()