This project was built in response to a real operational problem observed, where we needed a lightweight chatbot embedded into the website that answers customer questions using the company's FAQ and internal documents. We wanted: fast, accurate replies, an easy content ingest workflow for new PDFs, and complete control over the data (self-hosted where possible).
This repository contains a pragmatic RAG (retrieval-augmented generation) pipeline built around the following goals — a focused demo I built to add an FAQ-driven AI chatbot to a website:
- Accurate, context-aware answers driven by existing FAQ documents
- Simple document ingestion that anyone on the content team can use
- Automation for ingestion and chat logging using n8n
- Local vector store and privacy control (ChromaDB)
The pipeline described here is a real-world, minimal-ops solution that balances cost, reliability, and time-to-production.
The client had a lot of high-quality FAQ material and wanted a smarter UX on their website — not a simple keyword search but a conversational agent capable of short, precise answers and citations to the relevant FAQ sections.
Key requirements:
- Use the existing FAQ and knowledge documents (PDF) as the single source of truth.
- Host the knowledge and vector store in the client's infrastructure for privacy.
- Provide an automated ingestion process the content team can run.
- Capture chat transcripts for future improvements.
High-level architecture:
- LangChain + PDF loaders to extract and chunk text from PDFs.
- OpenAI-style embeddings to vectorize chunks (the repo expects local embedding support or remote service).
- ChromaDB as the vector store (self-hosted) for fast semantic similarity search.
- A lightweight query script
query.pythat computes embeddings and queries the store. - An ingestion script
add_data.pyused to add documents to ChromaDB and prevent duplicates. - n8n workflows (included in
n8n-workflows/) that automate ingestion, run queries from a chatbot, and log chats to Google Sheets.
This solution provides semantic retrieval (Chroma) to identify the most relevant text chunks and an LLM to produce helpful responses. The query.py + Code Tool pattern allows the agent (n8n or similar server) to use the vector store as a tool.
add_data.py— ingestion script. Use this to load new PDFs fromknowledge-base/into ChromaDB.query.py— run semantic queries against the cluster. When used by the chatbot it returns a JSON payload with scored segments.delete_collection.py— utility to remove a collection or clear the database.n8n-workflows/— exports for n8n: pipeline import, chatbot, and chat logger.knowledge-base/— PDF documents used as the base knowledge for the agent.
- A user asks a question through the web UI or messenger.
- n8n triggers the chatbot workflow and calls the
query.pytool (via the Code tool or Execute-Command node) to retrieve the top vector results. - The LLM (hosted or cloud) synthesizes a final answer using the retrieved chunks and formats a citation.
- Chat logs are appended to a Google Sheet for future review and training.
There are three example workflows included in n8n-workflows/:
chromadb-vector-RAG-Local-pipeline.json— sample ingestion pipeline that reads PDFs and ingests them into the local Chroma vector store.magnify-chatbot.json— the core chat agent. It uses the Code tool with an input schema to query the local vector store.knowledge-base/— PDF documents used as the base knowledge for the agent.
Before starting, ensure you have:
- Python 3.12+ installed
- ChromaDB server running on
localhost:8000(installation guide) - n8n (optional, for workflow automation) (installation guide)
- Basic familiarity with terminal/command line
- Start ChromaDB: follow the server docs (we assume it's on
localhost:8000). - Create and activate a Python environment, then install dependencies:
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt- Add documents to
knowledge-base/and run the ingestion script:
python add_data.py "_Magnify - Frequently Asked Questions (FAQ).pdf"- Run a quick query:
python query.py "How do I reset my account password?"- Import the n8n workflows to enable automatic ingestion and to run the chatbot from n8n.
Watch the chatbot in action:
Click above to see the live chatbot responding to customer questions using the FAQ knowledge base.
Example conversation showing FAQ-based responses
Chat interface integrated into the client's website
Vector search results and response generation
n8n chatbot workflow with AI agent and vector store integration
Document ingestion pipeline workflow
Key outcomes delivered
- ⚡ Response time: < 2 seconds average for FAQ queries
- 🎯 Accuracy: 90%+ relevant answers using vector similarity search
- 📉 Support reduction: Automated responses to common questions, reducing manual support load
- 🔄 Easy updates: Content team can add new PDFs without technical knowledge
- 📝 Chat logging: All conversations captured in Google Sheets for continuous improvement
- 🔒 Data privacy: Self-hosted vector store with full client control
Technical achievements:
- Seamless integration with existing website infrastructure
- Automated document ingestion pipeline via n8n
- Real-time semantic search across 10+ document chunks per query
- Production-ready chatbot delivered in under 4 weeks
- Demo accepted: The work was accepted, and the engagement closed. This repository stores the demo artifacts that powered the chatbot on the client's site.
- Import the
n8n-workflows/JSON exports into your n8n instance to enable the ingestion CLI and the chat agent. - Keep local venvs and the Chroma data directory out of version control using the provided
.gitignore. - If you accidentally pushed large files, use
git filter-repoorbfgto remove those from history — this repo already has a clean workflow branch available.
Problem: Error connecting to ChromaDB or connection refused
Solutions:
- Verify ChromaDB is running:
curl http://localhost:8000/api/v1/heartbeat - Check the port (default: 8000) and update scripts if using a different port
- Start ChromaDB server:
chroma run --path ./chroma_data --port 8000
Problem: ModuleNotFoundError: No module named 'chromadb' or similar
Solutions:
- Activate your virtual environment:
source .venv/bin/activate - Install dependencies:
pip install -r requirements.txt - Verify Python version:
python --version(requires 3.12+)
Problem: Queries return no results or empty collections
Solutions:
- Check collections exist: Run the collection check script or
curl http://localhost:8000/api/v1/collections - Ingest documents:
python add_data.py "your-document.pdf" - Verify PDF is in
knowledge-base/directory
Problem: Workflow fails or tools not recognized
Solutions:
- Ensure you're using the Code tool (not Function node) for vector search
- Verify the Execute Command node has the correct Python path
- Check n8n has access to the Python environment and scripts
- Check the ChromaDB docs
- Review n8n documentation
- Contact: Open an issue in this repository
Built with ❤️ by Samuel Kyere
Helping businesses automate with AI chatbots, content generation, and workflow integrations.
Thanks for reading — this is a demo version of the system we delivered to the client.