A RAG Chatbot with n8n and Pinecone Vector Database

Retrieval-Augmented Generation (RAG) allows Large Language Models (LLMs) to provide context-aware answers by retrieving information from an external vector database. In this post, we’ll walk through a complete n8n workflow that builds a chatbot capable of answering company policy questions using Pinecone Vector Database and OpenAI models.

Our setup has two main parts:

1. Data Loading to RAG – documents (company policies) are ingested from Google Drive, processed, embedded, and stored in Pinecone.
2. Data Retrieval using RAG – user queries are routed through an AI Agent that uses Pinecone to retrieve relevant information and generate precise answers.

1. Data Loading to RAG

This workflow section handles document ingestion. Whenever a new policy file is uploaded to Google Drive, it is automatically processed and indexed in Pinecone.

Nodes involved:

• Google Drive Trigger
  Watches a specific folder in Google Drive. Any new or updated file triggers the workflow.

• Google Drive (Download)
  Fetches the file (e.g., a PDF policy document) from Google Drive for processing.

• Recursive Character Text Splitter
  Splits long documents into smaller chunks (with a defined overlap). This ensures embeddings remain context-rich and retrieval works effectively.

• Default Data Loader
  Reads the binary document (PDF in this setup) and extracts the text.

• OpenAI Embeddings
  Generates high-dimensional vector representations of each text chunk using OpenAI’s embedding models.

• Pinecone Vector Store (Insert Mode)
  Stores the embeddings into a Pinecone index (n8ntest), under a chosen namespace. This step makes the policy data searchable by semantic similarity.

👉 Example flow: When HR uploads a new Work From Home Policy PDF to Google Drive, it is automatically split, embedded, and indexed in Pinecone.

2. Data Retrieval using RAG

Once documents are loaded into Pinecone, the chatbot is ready to handle user queries. This section of the workflow connects the chat interface, AI Agent, and retrieval pipeline.

Nodes involved:

• When Chat Message Received
  Acts as the webhook entry point when a user sends a question to the chatbot.

• AI Agent
  The core reasoning engine. It is configured with a system message instructing it to only use Pinecone-backed knowledge when answering.

• Simple Memory
  Keeps track of the conversation context, so the bot can handle multi-turn queries.

• Vector Store QnA Tool
  Queries Pinecone for the most relevant chunks related to the user’s question. In this workflow, it is configured to fetch company policy documents.

• Pinecone Vector Store (Query Mode)
  Acts as the connection to Pinecone, fetching embeddings that best match the query.

• OpenAI Chat Model
  Refines the retrieved chunks into a natural and concise answer. The model ensures answers remain grounded in the source material.

• Calculator Tool
  Optional helper if the query involves numerical reasoning (e.g., leave calculations or benefit amounts).

👉 Example flow: A user asks “How many work-from-home days are allowed per month?”. The AI Agent queries Pinecone through the Vector Store QnA tool, retrieves the relevant section of the HR policy, and returns a concise answer grounded in the actual document.

Wrapping Up

By combining n8n automation, Pinecone for vector storage, and OpenAI for embeddings + LLM reasoning, we’ve created a self-updating RAG chatbot.

• Data Loading pipeline ensures that every new company policy document uploaded to Google Drive is immediately available for semantic search.
• Data Retrieval pipeline allows employees to ask natural language questions and get document-backed answers.

This setup can easily be adapted for other domains — compliance manuals, tax regulations, legal contracts, or even product documentation.