Скачать или смотреть Building a Document QA Chatbot with n8n, Supabase, and Ollama by Blackcoffer Team

Objective
Set up an end-to-end pipeline using n8n on Windows to:
Ingest documents (e.g., from Google Drive),
Generate embeddings using Ollama,
Store embeddings in Supabase,
Retrieve relevant information from these embeddings,
Answer user queries using DeepSeek-R1 (OpenRouter).

Prerequisites
Windows 10/11
WSL (Windows Subsystem for Linux)
Git
Node.js (LTS)
Docker Desktop (for Ollama, if using Docker mode)
Supabase account
OpenRouter account with access to DeepSeek-R1

Architecture
The architecture illustrates an automated document processing QnA chatbot built using N8N. The workflow starts with a trigger that initiates file download from Google Drive. The downloaded document is loaded using a default data loader and split into chunks with a Recursive Character Text Splitter. These chunks are then converted into vector embeddings using the Ollama Embeddings model and stored in a Supabase Vector Store. When a user sends a message via the chat interface, it triggers the Question and Answer Chain. This chain uses a Vector Store Retriever to fetch relevant document embeddings from Supabase and processes the query with an OpenRouter chat model to generate a response. The system is designed to fully automate QnA by integrating document ingestion, embedding, storage, retrieval, and response generation.
Architecture



















Step 1: Set Up n8n on Windows
Option 1: Native (using Node.js)
npm install n8n -g
n8n

Option 2: Using Docker
docker run -it --rm \
-p 5678:5678 \
-v ~/.n8n:/home/node/.n8n \
n8nio/n8n

Access http://localhost:5678 in your browser.
Step 2: Set Up Ollama on Windows
Download Ollama
From the official site: https://ollama.com
Pull Model
ollama run all-minilm

This lightweight model is suitable for generating embeddings in your n8n pipeline.

Step 3: Create Supabase Vector Store
Create a Supabase Project
Create a table document_embeddings with columns:
id (uuid)
embedding (vector)
content (text)
metadata (jsonb)
created_at (timestamp)
Enable pgvector extension:
create extension if not exists vector;

Step 4: Workflow 1 - Document Embedding
Nodes Flow:
Trigger – Manual (when clicking “Test Workflow”)
Google Drive – Downloads selected .docx or .pdf
Default Data Loader – Extracts raw text
Recursive Character Text Splitter – Breaks into chunks
Embeddings (Ollama) – Generates embeddings
Supabase Vector Store – Stores them in document_embeddings

Step 5: Workflow 2 - Q&A Chatbot
Nodes Flow:
Trigger – On chat message received
Supabase Vector Store Retriever – Retrieves matching documents based on embedding similarity
OpenRouter Chat Model (DeepSeek-R1) – Generates human-like answers
Question and Answer Chain – Combines documents + user query for model
Response – Returns JSON like:
{ "response": "Your summarized answer..." }

Example in Action
User Prompt:
tell me project overview in short

Model Output:
The project involves integrating ChirpStack, InfluxDB, and Grafana... real-time monitoring dashboards... key sensors monitored include AM319, VS350...

Outcome
We have now a functioning document-based Q&A system using:
Documents from Google Drive
Embeddings via Ollama
Storage via Supabase
Responses from DeepSeek-R1 (OpenRouter)
This enables users to ask questions and get intelligent answers grounded in uploaded documents.