Wandbot is a question-answering bot designed specifically for Weights & Biases documentation. Leveraging the power of llama-index and OpenAI's gpt-4, it provides precise and context-aware responses using a combination of FAISS for RAG and OpenAI's gpt-4 for generating responses.

• Wandbot employs Retrieval Augmented Generation with a FAISS backend, ensuring efficient and accurate responses to user queries by retrieving relevant documents.
• It features periodic data ingestion and report generation, contributing to the bot's continuous improvement. You can view the latest data ingestion report here.
• The bot is integrated with Discord and Slack, facilitating seamless integration with these popular collaboration platforms.
• Performance monitoring and continuous improvement are made possible through logging and analysis with Weights & Biases Tables. Visit the workspace for more details here.
• Wandbot has a fallback mechanism for model selection, which is used when GPT-4 fails to generate a response.
• The bot's performance is evaluated using a mix of metrics, including retrieval accuracy, string similarity, and the correctness of model-generated responses.
• Curious about the custom system prompt used by the bot? You can view the full prompt here.

The project is built with Python version >=3.10.0,<3.11 and utilizes poetry for managing dependencies. Follow the steps below to install the necessary dependencies:

git clone [email protected]:wandb/wandbot.git
pip install poetry
cd wandbot
poetry install --all-extras
# Depending on which platform you want to run on run the following command:
# poetry install --extras discord # for discord
# poetry install --extras slack # for slack
# poetry install --extras api # for api

The data ingestion module pulls code and markdown from Weights & Biases repositories docodile and examples ingests them into vectorstores for the retrieval augmented generation pipeline. To ingest the data run the following command from the root of the repository

poetry run python -m src.wandbot.ingestion

You will notice that the data is ingested into the data/cache directory and stored in three different directories raw_data, vectorstore with individual files for each step of the ingestion process. These datasets are also stored as wandb artifacts in the project defined in the environment variable WANDB_PROJECT and can be accessed from the wandb dashboard.

Before running the Q&A bot, ensure the following environment variables are set:

OPENAI_API_KEY
COHERE_API_KEY
SLACK_EN_APP_TOKEN
SLACK_EN_BOT_TOKEN
SLACK_EN_SIGNING_SECRET
SLACK_JA_APP_TOKEN
SLACK_JA_BOT_TOKEN
SLACK_JA_SIGNING_SECRET
WANDB_API_KEY
DISCORD_BOT_TOKEN
COHERE_API_KEY
WANDBOT_API_URL="http://localhost:8000"
WANDB_TRACING_ENABLED="true"
WANDB_PROJECT="wandbot-dev"
WANDB_ENTITY="wandbot"

Once these environment variables are set, you can start the Q&A bot application using the following commands:

(poetry run uvicorn wandbot.api.app:app --host="0.0.0.0" --port=8000 > api.log 2>&1) & \
(poetry run python -m wandbot.apps.slack -l en > slack_en_app.log 2>&1) & \
(poetry run python -m wandbot.apps.slack -l ja > slack_ja_app.log 2>&1) & \
(poetry run python -m wandbot.apps.discord > discord_app.log 2>&1)

For more detailed instructions on installing and running the bot, please refer to the run.sh file located in the root of the repository.

Executing these commands will launch the API, Slackbot, and Discord bot applications, enabling you to interact with the bot and ask questions related to the Weights & Biases documentation.

We evaluated the performance of the Q&A bot manually and using auto eval strategies. The following W&B reports document the steps taken to evaluate the Q&A bot:

• How to evaluate an LLM Part 1: Building an Evaluation Dataset for our LLM System: The report dives into the steps taken to build a gold-standard evaluation set.
• How to evaluate an LLM Part 2: Manual Evaluation of our LLM System: The report talks about the thought process and steps taken to perform manual evaluation.
• How to evaluate an LLM Part 3: Auto-Evaluation; LLMs evaluating LLMs: Various LLM auto-eval startegies are documented in this report.

Manual Evaluation

We manually evaluated the Q&A bot's responses to establish a basline score.

Auto Evaluation (LLM evaluate LLM)

We employed a few auto evaluation strategies to speed up the iteration process of the bot's development

1. Creating Document Embeddings with FAISS
2. Constructing the Q&A Pipeline using llama-index
3. Selection of Models and Implementation of Fallback Mechanism
4. Deployment of the Q&A Bot on FastAPI, Discord, and Slack
5. Utilizing Weights & Biases Tables for Logging and Analysis
6. Evaluating the Performance of the Q&A Bot

You can monitor the usage of the bot in the following project: https://wandb.ai/wandbot/wandbot_public