A question answering bot for Weights & Biases documentation. This bot is built using llama-index and openai gpt-4.
- The bot utilizes retrieval augmented generation with FAISS backend to retrieve relevant documents and efficiently handle user queries and provides accurate, context-aware responses
- Periodic data ingestion with report generation for continuous improvement of the bot.: Checkout the latest data ingestion report here
- Integrated with Discord and Slack, allowing seamless integration into popular collaboration platforms.
- Logging and analysis with Weights & Biases Tables for performance monitoring and continuous improvement.: Checkout the workspace for more details here
- Uses a fallback mechanism for model selection when GPT-4 is unable to generate a response.
- Evaluation using a combination of metrics such as retrieval accuracy, string similarity, and model-generated response correctness
- Want to know more about the custom system prompt used by the bot?: Checkout the full prompt here
The project uses python = ">=3.10.0,<3.11" and uses poetry for dependency management. To install the dependencies:
git clone [email protected]:wandb/wandbot.git
pip install poetry --all-extras
cd wandbot
poetry install
# 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 apiThe 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.ingestionYou 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.
You will need to set the following environment variables:
OPENAI_API_KEY
SLACK_APP_TOKEN
SLACK_BOT_TOKEN
SLACK_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"Then you can run the Q&A bot application, use 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 > slack_app.log 2>&1) & \
(poetry run python -m wandbot.apps.discord > discord_app.log 2>&1)Please refer to the run.sh file in the root of the repository for more details on commands related to installing and running the bot.
This will start the chatbot applications - the api, the slackbot and the discord bot, allowing you to interact with it and ask questions related to the Weights & Biases documentation.
To evaluate the performance of the Q&A bot, the provided evaluation script (…) can be used. This script utilizes a separate dataset for evaluation, which can be stored as a W&B Artifact. The evaluation script calculates retrieval accuracy, average string distance, and chat model accuracy.
The evaluation script downloads the evaluation dataset from the specified W&B Artifact, performs the evaluation using the Q&A bot, and then logs the results, such as retrieval accuracy, average string distance, and chat model accuracy, back to W&B. The logged results can be viewed on the W&B dashboard.
To run the evaluation script, use the following commands:
cd wandbot
poetry run python -m eval- Document Embeddings with FAISS
- Building the Q&A Pipeline with llama-index
- Model Selection and Fallback Mechanism
- Deploying the Q&A Bot on FastAPI, Discord and Slack
- Logging and Analysis with Weights & Biases Tables
- Evaluation of the Q&A Bot
You can track the bot usage in the following project: https://wandb.ai/wandbot/wandbot_public