RAG-based n8n Intelligent Q&A Bot: Practice with LangChain and Vector Databases

This project is an academic AI assignment implementation that shows how to build a professional FAQ bot for the n8n open-source workflow automation platform using the Retrieval-Augmented Generation (RAG) architecture. The core tech stack includes the LangChain framework, ChromaDB vector database, sentence-transformers/all-MiniLM-L6-v2 embedding model, and Hugging Face-hosted LLM services. By combining information retrieval and text generation, the project achieves fact-based accurate answers while allowing traceability of answer sources and avoiding model hallucinations.

Retrieval-Augmented Generation (RAG) is a technology that combines information retrieval and text generation, solving the problem of traditional Q&A systems either relying on the model's internal knowledge (prone to hallucinations) or only returning document fragments (lacking integration capabilities). Its advantages include accurate answers (based on real documents), ability to handle new knowledge outside training data, traceable answers, and avoidance of content inconsistent with documents. The project chose n8n as the target product because it is a popular open-source workflow automation platform, widely used in AI automation, system integration, and other fields, making it suitable as a knowledge base source for the RAG system.

The project architecture consists of three core modules:

1. Dataset Construction Module (build_dataset.py)：Crawl n8n official documentation pages via web crawlers, extract text content, and save it as a CSV-format dataset, which serves as the data foundation for the RAG process.
2. Vector Embedding and Storage Module (ingest.py)：Split long documents into small fragments suitable for retrieval; generate text vectors using the sentence-transformers/all-MiniLM-L6-v2 model; store vectors and corresponding text fragments in ChromaDB (an open-source embedded vector database that requires no additional infrastructure).
3. Q&A Interaction Module (chatbot.py)：When a user asks a question, the system converts the question into a vector, retrieves the most similar document fragments from ChromaDB, constructs a prompt by combining the question and fragments, calls the Hugging Face-hosted LLM to generate an answer, and returns the answer along with source links. In terms of technology selection: The LangChain framework handles document loading and splitting, encapsulates embedding models and vector storage, and builds the retrieval-generation execution chain; the all-MiniLM-L6-v2 model was chosen as the embedding model due to its small size, fast inference, and open-source nature, which is suitable for academic scenarios; the Hugging Face-hosted inference service reduces hardware requirements, requiring only an access token to use.

The implementation is divided into three phases:

1. Data Preparation Phase：Crawl specific pages of n8n official documentation, extract Q&A content, and organize it into a structured CSV dataset covering n8n's basic introduction, AI workflow support, AI Agent tool usage, common questions, etc.
2. Vector Database Construction Phase：Run ingest.py to load the CSV dataset, split documents using the LangChain text splitter, generate vectors for each fragment, and store them in ChromaDB (only needs to be executed once unless the knowledge base is updated).
3. Q&A Service Phase：Launch chatbot.py; when a user inputs a natural language question (e.g., "What is n8n?"), the system retrieves relevant fragments in real time to generate a fact-based answer; if no relevant information is found, it returns a fallback response to avoid false content.

As an academic assignment in AI and natural language processing, this project covers multiple important concepts and technologies: semantic search and vector retrieval (mapping text to semantic space via embedding models), RAG architecture (combining retrieval and generation to build reliable Q&A systems), LLM integration (calling hosted LLM services), vector database applications (mastering ChromaDB usage), and Python AI development ecosystem (familiarizing with mainstream libraries like LangChain and Transformers).

The limitations of the current implementation include: limited knowledge coverage (only answers content from selected documents), retrieval quality dependent on chunking strategy, and hosted model output quality affected by service status. Future improvement directions: expand the knowledge base to cover more n8n documents, try different embedding models (e.g., OpenAI text-embedding-3), implement multi-turn dialogue support, and add user feedback mechanisms to optimize retrieval effectiveness.