FAQ Slack Bot for DataTalks.Club Zoomcamps

An LLM-powered QA Slack chatbot that answers questions related to DataTalks.Club Zoomcamps (ML, DE, MLOps, and LLM courses). The bot uses RAG (Retrieval Augmented Generation) with vector search over indexed course materials to provide accurate, context-aware answers to student questions.

What Does This Bot Do?

The bot automatically responds to questions in Slack channels dedicated to DataTalks.Club courses by:

• Searching indexed course materials including FAQ documents, Slack history, GitHub repositories, and YouTube subtitles
• Using vector similarity search to find the most relevant content for each question
• Re-ranking results with Cohere's re-ranking model to improve relevance
• Generating contextual answers using GPT-4o-mini based on retrieved course materials
• Providing source references with links back to original materials (Slack messages, docs, GitHub files, videos)
• Learning from feedback through upvote/downvote reactions tracked via LangSmith
• Prioritizing recent content with time-weighted ranking to surface up-to-date information

The bot maintains separate query engines for each Zoomcamp and routes questions based on the Slack channel ID.

For a detailed explanation of how the bot works, including architecture and evaluation results, see this comprehensive report.

Architecture Overview

The system consists of two main components:

1. Slack Bot (slack_bot/)

• Runs as a Slack app using Socket Mode with the Slack Bolt framework
• Maintains 4 separate query engines (ML, DE, MLOps, LLM Zoomcamps)
• Uses LlamaIndex with Milvus/Zilliz vector store and HuggingFace embeddings
• Implements time-weighted post-processing to prioritize recent Slack messages (applied first)
• Implements Cohere re-ranking to improve result relevance (20 retrieved → 10 after time-weighting → 4 after re-ranking)

2. Ingestion Pipeline (ingest/)

• Indexes data from multiple sources: GitHub FAQs, Slack history, GitHub repos, YouTube subtitles
• Uses embeddings cache (Upstash Redis) to reduce costs
• Runs on schedule via GitHub Actions or can be triggered manually
• Supports both local (Milvus) and production (Zilliz Cloud) vector stores

Getting Started

Prerequisites

• Python 3.10
• Conda (recommended) or virtualenv
• Docker and Docker Compose (for local Milvus instance)
• Slack workspace with bot permissions
• Required API keys (see Environment Variables section)

Quick Start

1. Running the Bot Locally

1. Set up the environment:

   conda create --name slack-bot python=3.10
   conda activate slack-bot
   cd slack_bot
   pip install -r requirements.txt

2. Configure environment variables:

   • In the slack_bot/ directory, rename dev.env to .env
   • Fill in all required values (see Environment Variables)

3. Start local Milvus instance (in a separate terminal):

   cd ingest/local_milvus
   docker compose up

4. Run ingestion to populate the index (in another terminal)

   Important: Before running the bot, you must populate the vector store with course data. See the Running Ingestion Locally section below for detailed instructions.

5. Start the bot (in the original terminal or a new one):

   cd slack_bot
   source .env
   python main.py

For more details, see slack_bot/README.md.

2. Running Ingestion Locally

1. Start local Milvus (in a separate terminal):

   cd ingest/local_milvus
   docker compose up

2. Configure environment variables:

   • In the ingest/ directory, rename dev.env to .env
   • Fill in all required values

3. Run ingestion for specific courses (in another terminal, from the project root):

   source ingest/.env
   export PYTHONPATH="${PYTHONPATH}:$(pwd)"
   python ingest/ml/ingest_ml.py      # for ML Zoomcamp
   python ingest/de/ingest_de.py      # for DE Zoomcamp
   python ingest/llm/ingest_llm.py    # for LLM Zoomcamp
   python ingest/mlops/ingest_mlops.py # for MLOps Zoomcamp

For more details, see ingest/local_development.md.

Environment Variables

⚠️ Security Note: Never commit .env files or expose API keys in your code. Keep all credentials private and ensure .env is listed in your .gitignore.

Required for Slack Bot

• SLACK_BOT_TOKEN - Bot token for Slack app
• SLACK_APP_TOKEN - App-level token for Socket Mode
• OPENAI_API_KEY - OpenAI API key for GPT-4o-mini
• COHERE_API_KEY - Cohere API key for result re-ranking
• ZILLIZ_CLOUD_URI, ZILLIZ_CLOUD_API_KEY - For ML/DE Zoomcamps
• ZILLIZ_PUBLIC_ENDPOINT, ZILLIZ_API_KEY - For MLOps/LLM Zoomcamps
• LANGCHAIN_API_KEY, LANGCHAIN_PROJECT - For LangSmith logging
• LOCAL_MILVUS - Set to true for local development

Required for Ingestion

• All of the above, plus:
• GH_TOKEN - GitHub personal access token for API access
• UPSTASH_REDIS_REST_URL, UPSTASH_REDIS_REST_TOKEN - For embeddings cache
• EMBEDDING_CACHE_NAMESPACE - Set per course to avoid cache collisions
• EXECUTION_ENV - Set to local for local dev, zilliz-cluster for MLOps/LLM production, or leave empty for ML/DE production

Deployment

Slack Bot

• Deployed to Fly.io
• Docker image: aaalexlit/faq-slack-bot
• Automated via GitHub Actions:
  • .github/workflows/docker-image.yml - Builds Docker image
  • .github/workflows/fly-deploy.yml - Deploys to Fly.io

Ingestion Pipeline

• Runs on schedule via GitHub Actions (see .github/workflows/schedule-ingest-*.yml)
• Docker image: aaalexlit/zoomcamp-faq-ingest
• Can be triggered manually via .github/workflows/run-ingest.yml

Technology Stack

• Programming Language: Python 3.10
• LLM Framework: LlamaIndex
• Vector Database: Milvus (local), Zilliz Cloud (production)
• Embeddings: HuggingFace BAAI/bge-base-en-v1.5
• Re-ranking: Cohere Rerank API
• LLM: OpenAI GPT-4o-mini
• Slack Integration: Slack Bolt (Socket Mode)
• Observability: LangSmith
• Embeddings Cache: Upstash Redis

Project Structure

├── slack_bot/          # Slack bot application
│   ├── main.py         # Entry point
│   └── README.md       # Bot setup instructions
├── ingest/             # Data ingestion pipeline
│   ├── ml/             # ML Zoomcamp ingestion
│   ├── de/             # DE Zoomcamp ingestion
│   ├── llm/            # LLM Zoomcamp ingestion
│   ├── mlops/          # MLOps Zoomcamp ingestion
│   ├── readers/        # Custom data readers
│   ├── utils/          # Shared utilities
│   └── local_milvus/   # Docker Compose for local Milvus
├── .github/workflows/  # CI/CD workflows
└── CLAUDE.md           # Development guidelines for Claude Code

Contributing

For development guidelines and detailed implementation notes, see CLAUDE.md.

License

This project is part of DataTalks.Club community resources.