argilla-sdk-chatbot

Building a chatbot for Argilla SDK Step by step.

This repository contains all the scripts and notebooks that were used to create a ChatBot to help with the new Argilla 2.0 SDK.

The blog is at argilla-chatbot.md file. Will share the link to the deployed version once accepted at https://github.com/huggingface/blog.

The blogpost gives a tour of the different components needed to create a chatbot (in this case for the Argilla 2.0 SDK, but can be applied to any other repository) just using the raw documentation in a repository, and interact with the final app deployed at a Hugging Face Space.

To interact with the chatbot, you can visit the following link:

https://huggingface.co/spaces/plaguss/argilla-sdk-chatbot-space

Contents

Each section contains a brief explanation of what the script/notebook does, how to run it and what artifacts should be obtained from it.

1. docs_dataset.py

The first script to run. It goes through a repo docs folder (must be in markdown), creates chunks from it and creates a dataset.

Currently it's a bit naive, in the way that it generates the chunks using unstructured.io's chunk_by_title (with the default values from the function, nothing else was tweaked).

It's a really short script to showcase the first step.

Needs the requirements from requirements/requirements_docs_dataset.txt.

Run:

python docs_dataset.py \
    "argilla-io/argilla-python" \
    --dataset-name "plaguss/argilla_sdk_docs_raw_unstructured"

It will generate the following dataset: plaguss/argilla_sdk_docs_raw_unstructured.

2. pipeline_docs_queries.py

Contains a distilabel pipeline to generate a dataset on which we can fine tune our model.

Needs distilabel. Uses the Hugging Face's Inference Endpoints.

Run it as:

python docs_dataset.py \
    "argilla-io/argilla-python" \
    --dataset-name "plaguss/argilla_sdk_docs_raw_unstructured"

Will generate a dataset like the following: plaguss/argilla_sdk_docs_queries.

3. argilla_datasets.ipynb

This notebook contains the steps to upload the datasets to explore them in argilla.

The Hugging Face Space contains the datasets: Argilla datasets. They use the same names as in the Hugging Face Hub.

4. train_embeddings.ipynb

It follows this blog from Phillip Schmid to fine tune a model on the dataset we previously generated.

Needs the requirements from requirements/requirements_training.txt.

Final model: plaguss/bge-base-argilla-sdk-matryoshka.

The model doesn't improve that much as can be seen in the notebook, but does the job.

5. vector_db.ipynb

Creates the vector database. Uses lancedb, it's a small embedded database that can be easily moved.

The database is a file stored with the distiset, the file lancedb.tar.gz. This file can be moved as any other artifact, for example download it in the same place the app will be deployed.

6. app/

Contains the gradio app with the chatbot.

It's a basich Chat app with gradio. Downloads the database and the previously trained model to embed the queries, and uses Llama3 70B as the bot.

The prompt is heavily inspired from that of wandbot.

To run it locally.

python app/app.py

Needs more guardrails, but some initial examples can be seen already:

(still need to remove the eot_id)

• One of the queries from the examples:

• An unrelated question:

• Ask it to ignore the system prompt:

TODO

• The dataset should be rebuilt with an updated version of the docs?: updated docs

•  Update code to use client.chat_completion, for reference: argilla-io/distilabel#772.

Next steps

• Include references/sources for the responses, so we can give some URL to visit.