Build a Website RAG Chatbot with Gemini

How It Works

Large language models only know what they were trained on, so a chatbot that answers questions about your business needs an outside source of truth. This n8n template wires that source of truth to Gemini using a retrieval-augmented generation pattern: a scheduled job scrapes your website with the Apify Website Content Crawler, pushes the text into a Pinecone vector index, and a separate chat agent retrieves the most relevant chunks before each response. The workflow is split into two sub-logics, flagged in the canvas with green sticky notes, so the ingestion pipeline and the live agent can be edited independently.

Setup Steps

1. Create a Google Cloud project: Enable the Vertex AI API and generate a Google AI API key from Google AI Studio.
2. Create an Apify account: Sign up at apify.com; the Apify node uses the Website Content Crawler actor to pull every page from your site.
3. Create a Pinecone index: On the Pinecone free tier, create an index named company-website and copy the API key from the dashboard.
4. Add credentials in n8n: Configure the Google Gemini (PaLM) API and Pinecone API credentials so both the embedding and vector store nodes can authenticate.
5. Set the schedule frequency: Edit the Schedule Trigger to match how often your site content changes; swap it for a click trigger if the chatbot only needs to be trained once.
6. Configure the Apify node: Authenticate the Apify node (OAuth or API token) and set the target website URL inside the JSON input before the first run.

Automation Workflow

Schedule Trigger → Apify Website Content Crawler → Embeddings Gemini → Pinecone Vector Store (insert)
                                                                            ↓
Chat Trigger → AI Agent → Vector Store Tool → Pinecone Vector Store (query)
                   ↓
          Google Gemini Chat Model (gemini-pro) → Chat response

Benefits

• Grounded answers: Every reply is built from the exact text chunks retrieved from Pinecone, so the chatbot cites what your site actually says rather than guessing.
• Fresh index: The Schedule Trigger reruns the crawl on your cadence, so content changes on the website propagate into the vector store without manual work.
• Model portability: The Gemini chat node can be swapped for another LLM; only the embedding step needs to stay aligned with the index.
• Reusable across sites: The Website Content Crawler accepts any URL, so the same template can index a competitor site, a help center, or a mix of sources.
• Clear separation: Ingestion and chat live in two distinct sub-workflows, which makes it easy to pause one while editing the other.

Advanced Features

• Two independent sub-workflows for ingestion and chat, sharing only the Pinecone index.
• Scheduled re-crawl that updates embeddings for pages that have changed since the last run.
• AI Agent node that chains a Vector Store Tool with the Gemini chat model, so retrieval and generation are orchestrated inside one agent step.
• JSON-driven Apify input, which lets you extend the crawl to multiple start URLs or combine several websites into one knowledge base.

Use Cases

• Answer product questions on a marketing site using the latest copy from the product pages.
• Build an internal help desk bot that stays in sync with a documentation portal.
• Give sales reps a chat interface over a competitor’s public knowledge base.
• Power an onboarding assistant that pulls policies and procedures from an intranet.
• Combine several domains into one chatbot so visitors can ask about parent and subsidiary brands in a single thread.