n8n template · RAG Assistant Pack

Build a RAG chatbot with citations in n8n

The heart of the RAG Assistant Pack: a public n8n chat that queries your Supabase vector store and answers with source citations. Six nodes — proof that production RAG requires neither a framework nor code.

$ n8n import:workflow --input=chatbot-rag-citations.json
✔ Import successful — 6 nodes, valid connections

The problem it solves

A raw LLM answers everything, including wrongly: ask it your refund policy and it will invent a plausible one. For business use, the answer must come from your documents, and be verifiable.

RAG solves both requirements: vector search supplies the relevant passages, and the prompt constrains the model to claim nothing outside that context and to cite every source. Users can verify; trust follows.

How the workflow works, node by node

Node names are in French, exactly as they appear in the imported workflow — the logic is language-agnostic.

1. 01

   Chat

   Chat (trigger)

   Chat interface hosted by n8n, reachable by public URL — shareable with the team without building a frontend. Also manages the session ID.

2. 02

   Assistant documentaire

   RAG chain (Q&A)

   Retrieval Q&A chain: it injects retrieved passages into the prompt ({context}) and requires answering only from them, with [source] citations and a polite refusal out of scope.

3. 03

   Modèle OpenAI

   OpenAI model

   ai_languageModel sub-node (gpt-4o-mini). Move up to a stronger model if your documents are dense or questions span many documents.

4. 04

   Retriever

   Retriever

   ai_retriever sub-node: brings back the 4 chunks closest to the question (adjustable topK — more context versus more noise).

5. 05

   Base vectorielle Supabase

   Supabase vector store

   Retrieve mode on the documents table: runs the similarity search through the match_documents function, Postgres-side.

6. 06

   Embeddings OpenAI

   OpenAI embeddings

   Vectorizes the user's question with the same model as ingestion (text-embedding-3-small) — essential: question and documents must live in the same vector space.

What you need to run it

• ✔n8n ≥ 1.60 with the LangChain nodes
• ✔A Supabase pgvector store already populated (PDF or Notion ingestion workflows from the same pack)
• ✔An OpenAI API key
• ✔The match_documents SQL function installed (provided in the pack guide)

Customization ideas

• →Adapt the prompt's tone and language (answer in English even over French documents, for instance)
• →Raise topK to 6–8 for cross-cutting questions, at a slightly higher token cost
• →Restrict access by disabling the chat's public mode and integrating it into your internal tool via the pack's API workflow