app.py

import gradio as gr
from gradio_huggingfacehub_search import HuggingfaceHubSearch
import nbformat as nbf
from huggingface_hub import HfApi
from httpx import Client
import logging
from huggingface_hub import InferenceClient
import json
import re
import pandas as pd
from gradio.data_classes import FileData
from utils.prompts import (
    generate_mapping_prompt,
    generate_user_prompt,
    generate_rag_system_prompt,
    generate_eda_system_prompt,
)

"""
TODOs:
- Need feedback on the output commands to validate if operations are appropiate to data types 
- Refactor
- Make the notebook generation more dynamic, add loading components to do not freeze the UI
- Fix errors:
    - When generating output
    - When parsing output
    - When pushing notebook
- Add target tasks to choose for the notebook:
    - Exploratory data analysis
    - Auto training
    - RAG
    - etc.
- Enable 'generate notebook' button only if dataset is available and supports library
- First get compatible-libraries and let user choose the library
"""

# Configuration
BASE_DATASETS_SERVER_URL = "https://datasets-server.huggingface.co"
HEADERS = {"Accept": "application/json", "Content-Type": "application/json"}

client = Client(headers=HEADERS)
inference_client = InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct")

logging.basicConfig(level=logging.INFO)


def get_compatible_libraries(dataset: str):
    try:
        response = client.get(
            f"{BASE_DATASETS_SERVER_URL}/compatible-libraries?dataset={dataset}"
        )
        response.raise_for_status()
        return response.json()
    except Exception as e:
        logging.error(f"Error fetching compatible libraries: {e}")
        raise


def create_notebook_file(cell_commands, notebook_name):
    nb = nbf.v4.new_notebook()
    nb["cells"] = [
        nbf.v4.new_code_cell(
            cmd["source"]
            if isinstance(cmd["source"], str)
            else "\n".join(cmd["source"])
        )
        if cmd["cell_type"] == "code"
        else nbf.v4.new_markdown_cell(cmd["source"])
        for cmd in cell_commands
    ]

    with open(notebook_name, "w") as f:
        nbf.write(nb, f)
    logging.info(f"Notebook {notebook_name} created successfully")


def get_first_rows_as_df(dataset: str, config: str, split: str, limit: int):
    try:
        resp = client.get(
            f"{BASE_DATASETS_SERVER_URL}/first-rows?dataset={dataset}&config={config}&split={split}"
        )
        resp.raise_for_status()
        content = resp.json()
        rows = content["rows"]
        rows = [row["row"] for row in rows]
        first_rows_df = pd.DataFrame.from_dict(rows).sample(frac=1).head(limit)
        features = content["features"]
        features_dict = {feature["name"]: feature["type"] for feature in features}
        return features_dict, first_rows_df
    except Exception as e:
        logging.error(f"Error fetching first rows: {e}")
        raise


def get_txt_from_output(output):
    try:
        extracted_text = extract_content_from_output(output)
        content = json.loads(extracted_text)
        logging.info(content)
        return content
    except Exception as e:
        gr.Error("Error when parsing notebook, try again.")
        logging.error(f"Failed to fetch compatible libraries: {e}")
        raise


def extract_content_from_output(output):
    patterns = [r"`json(.*?)`", r"```(.*?)```"]

    for pattern in patterns:
        match = re.search(pattern, output, re.DOTALL)
        if match:
            return match.group(1)

    try:
        index = output.index("```json")
        logging.info(f"Index: {index}")
        return output[index + 7 :]
    except ValueError:
        logging.error("Unable to generate Jupyter notebook.")
        raise


def content_from_output(output):
    pattern = r"`json(.*?)`"
    match = re.search(pattern, output, re.DOTALL)
    if not match:
        pattern = r"```(.*?)```"
        match = re.search(pattern, output, re.DOTALL)
        if not match:
            try:
                index = output.index("```json")
                logging.info(f"Index: {index}")
                return output[index + 7 :]
            except:
                pass
            raise Exception("Unable to generate jupyter notebook.")
    return match.group(1)


def generate_eda_cells(dataset_id, profile: gr.OAuthProfile | None):
    for messages in generate_cells(dataset_id, generate_eda_system_prompt, "eda"):
        yield messages, gr.update(visible=False), None  # Keep button hidden

    yield (
        messages,
        gr.update(visible=profile and dataset_id.split("/")[0] == profile.username),
        f"{dataset_id.replace('/', '-')}-eda.ipynb",
    )


def generate_rag_cells(dataset_id, profile: gr.OAuthProfile | None):
    for messages in generate_cells(dataset_id, generate_rag_system_prompt, "rag"):
        yield messages, gr.update(visible=False), None  # Keep button hidden

    yield (
        messages,
        gr.update(visible=profile and dataset_id.split("/")[0] == profile.username),
        f"{dataset_id.replace('/', '-')}-rag.ipynb",
    )


def generate_embedding_cells(dataset_id, profile: gr.OAuthProfile | None):
    for messages in generate_cells(dataset_id, generate_embedding_prompt, "embedding"):
        yield messages, gr.update(visible=False), None  # Keep button hidden

    yield (
        messages,
        gr.update(visible=profile and dataset_id.split("/")[0] == profile.username),
        f"{dataset_id.replace('/', '-')}-embedding.ipynb",
    )


def push_to_hub(
    history,
    dataset_id,
    notebook_file,
    profile: gr.OAuthProfile | None,
    oauth_token: gr.OAuthToken | None,
):
    logging.info(f"Pushing notebook to hub: {dataset_id} on file {notebook_file}")
    if not profile or not oauth_token:
        yield history + [
            gr.ChatMessage(role="assistant", content="⏳ _Login to push to hub..._")
        ]
        return
    logging.info(f"Profile: {profile}, token: {oauth_token.token}")

    notebook_name = "dataset_analysis.ipynb"
    api = HfApi(token=oauth_token.token)
    try:
        logging.info(f"About to push {notebook_file} - {notebook_name} - {dataset_id}")
        api.upload_file(
            path_or_fileobj=notebook_file,
            path_in_repo=notebook_name,
            repo_id=dataset_id,
            repo_type="dataset",
        )
        link = f"https://huggingface.co/datasets/{dataset_id}/blob/main/{notebook_name}"
        logging.info(f"Notebook pushed to hub: {link}")
        yield history + [
            gr.ChatMessage(
                role="user",
                content=f"[See the notebook on the Hub]({link})",
            )
        ]
    except Exception as e:
        logging.info("Failed to push notebook", e)
        yield history + [gr.ChatMessage(role="assistant", content=e)]


def generate_cells(dataset_id, prompt_fn, notebook_type="eda"):
    try:
        libraries = get_compatible_libraries(dataset_id)
    except Exception as err:
        gr.Error("Unable to retrieve dataset info from HF Hub.")
        logging.error(f"Failed to fetch compatible libraries: {err}")
        return []

    if not libraries:
        gr.Error("Dataset not compatible with pandas library.")
        logging.error(f"Dataset not compatible with pandas library")
        return gr.File(visible=False), gr.Row.update(visible=False)

    pandas_library = next(
        (lib for lib in libraries.get("libraries", []) if lib["library"] == "pandas"),
        None,
    )
    if not pandas_library:
        gr.Error("Dataset not compatible with pandas library.")
        return []

    first_config_loading_code = pandas_library["loading_codes"][0]
    first_code = first_config_loading_code["code"]
    first_config = first_config_loading_code["config_name"]
    first_split = list(first_config_loading_code["arguments"]["splits"].keys())[0]
    features, df = get_first_rows_as_df(dataset_id, first_config, first_split, 3)
    prompt = generate_user_prompt(
        features, df.head(5).to_dict(orient="records"), first_code
    )
    messages = [gr.ChatMessage(role="user", content=prompt)]
    yield messages + [gr.ChatMessage(role="assistant", content="⏳ _Starting task..._")]

    prompt_messages = [
        {"role": "system", "content": prompt_fn()},
        {"role": "user", "content": prompt},
    ]
    output = inference_client.chat_completion(
        messages=prompt_messages, stream=True, max_tokens=2500
    )

    generated_text = ""
    current_line = ""
    for chunk in output:
        current_line += chunk.choices[0].delta.content
        if current_line.endswith("\n"):
            generated_text += current_line
            messages.append(gr.ChatMessage(role="assistant", content=current_line))
            current_line = ""
        yield messages
    yield messages

    logging.info("---> Formated prompt")
    formatted_prompt = generate_mapping_prompt(generated_text)
    logging.info(formatted_prompt)
    prompt_messages = [{"role": "user", "content": formatted_prompt}]
    yield messages + [
        gr.ChatMessage(role="assistant", content="⏳ _Generating notebook..._")
    ]

    output = inference_client.chat_completion(
        messages=prompt_messages, stream=False, max_tokens=2500
    )
    cells_txt = output.choices[0].message.content
    logging.info("---> Model output")
    logging.info(cells_txt)

    commands = get_txt_from_output(cells_txt)
    html_code = f"<iframe src='https://huggingface.co/datasets/{dataset_id}/embed/viewer' width='80%' height='560px'></iframe>"

    commands.insert(
        0,
        {
            "cell_type": "code",
            "source": f'from IPython.display import HTML\n\ndisplay(HTML("{html_code}"))',
        },
    )
    commands.insert(0, {"cell_type": "markdown", "source": "# Dataset Viewer"})
    notebook_name = f"{dataset_id.replace('/', '-')}-{notebook_type}.ipynb"
    create_notebook_file(commands, notebook_name=notebook_name)
    messages.append(
        gr.ChatMessage(role="user", content="Here is the generated notebook file")
    )
    yield messages
    messages.append(
        gr.ChatMessage(
            role="user",
            content=FileData(path=notebook_name, mime_type="application/x-ipynb+json"),
        )
    )
    yield messages


def coming_soon_message():
    return gr.Info("Coming soon")


with gr.Blocks(fill_height=True) as demo:
    gr.Markdown("# 🤖 Dataset notebook creator 🕵️")
    with gr.Row():
        with gr.Column(scale=1):
            dataset_name = HuggingfaceHubSearch(
                label="Hub Dataset ID",
                placeholder="Search for dataset id on Huggingface",
                search_type="dataset",
                value="",
            )

            @gr.render(inputs=dataset_name)
            def embed(name):
                if not name:
                    return gr.Markdown("### No dataset provided")
                html_code = f"""
                <iframe
                src="https://huggingface.co/datasets/{name}/embed/viewer/default/train"
                frameborder="0"
                width="100%"
                height="350px"
                ></iframe>
                """
                return gr.HTML(value=html_code)

            with gr.Row():
                generate_eda_btn = gr.Button("Generate EDA notebook")
                generate_embedding_btn = gr.Button("Generate Embeddings notebook")
                generate_rag_btn = gr.Button("Generate RAG notebook")
                generate_training_btn = gr.Button("Generate Training notebook")
        with gr.Column():
            chatbot = gr.Chatbot(
                label="Results",
                type="messages",
                avatar_images=(
                    None,
                    None,
                ),
            )
            with gr.Row():
                login_btn = gr.LoginButton()
                push_btn = gr.Button("Push to hub", visible=False)
    notebook_file = gr.File(visible=False)
    generate_eda_btn.click(
        generate_eda_cells,
        inputs=[dataset_name],
        outputs=[chatbot, push_btn, notebook_file],
    )

    generate_rag_btn.click(
        generate_rag_cells,
        inputs=[dataset_name],
        outputs=[chatbot, push_btn, notebook_file],
    )

    generate_embedding_btn.click(
        generate_embedding_cells,
        inputs=[dataset_name],
        outputs=[chatbot, push_btn, notebook_file],
    )

    generate_training_btn.click(coming_soon_message, inputs=[], outputs=[])
    push_btn.click(
        push_to_hub,
        inputs=[
            chatbot,
            dataset_name,
            notebook_file,
        ],
        outputs=[chatbot],
    )

demo.launch()