n8n Template: WhatsApp AI Chatbot (Text/Voice/Image/Document + RAG)

Auto-index Google Docs product documentation into MongoDB Atlas, receive multi-type WhatsApp messages, and answer with GPT-4o-mini using retrieval-augmented generation

Who

Internal support teams, product specialists, and knowledge managers

You get

Users send text, audio, image, or document via WhatsApp; the system retrieves product docs and returns context-aware AI answers

You need

n8n · mongodb-atlas · whatsapp · google-docs

Steps

1. Configure vector embeddings and document indexing environment

   Set up connections for the ingestion workflow: 1) Authorize Google Docs OAuth2 and enter the product document URL; 2) Connect MongoDB Atlas cluster in the MongoDB node credentials, specify collection name (e.g., n8n-template); 3) Ensure a vector search index exists on the collection with the same name configured in n8n (default data_index); 4) Create the index using the MongoDB search index template below.

   Setup guidance · Source text。MongoDB Atlas search index JSON definition; create via Atlas console or CLI

   Copy prompt

   {
     "mappings": {
       "dynamic": false,
       "fields": {
         "_id": { "type": "string" },
         "text": { "type": "string" },
         "embedding": {
           "type": "knnVector",
           "dimensions": 1536,
           "similarity": "cosine"
         },
         "source": { "type": "string" },
         "doc_id": { "type": "string" }
       }
     }
   }

2. Configure WhatsApp Q&A workflow

   Configure connections for receiving and replying to WhatsApp messages: 1) Enter Meta account credentials in the WhatsApp Trigger and WhatsApp nodes to enable message receiving and sending; 2) Connect the MongoDB Vector Search node to the same collection (n8n-template) with index name data_index; 3) Set the system prompt in the Knowledge Base Agent node to reflect your company's tone, answering style, and business rules, ensuring it references the connected MongoDB collection for context retrieval.

3. Execute document ingestion and indexing workflow

   Manually trigger 'Execute Workflow' to run the ingestion workflow. Flow: Google Docs Importer fetches document → Document Section Loader loads content with doc_id metadata → Document Chunker splits by markdown (chunkSize=3000, chunkOverlap=200) → OpenAI Embeddings Generator creates 1536-dim vectors → MongoDB Vector Store Inserter writes chunks and embeddings to Atlas. After successful run, the product documentation is indexed for semantic search.

   Setup guidance · Source text。Sticky Note instruction; reminder before manual workflow trigger

   Copy prompt

   Run this workflow manually to import and index Google Docs product documentation into MongoDB with vector embeddings for fast search.

4. Run WhatsApp AI Q&A workflow

   The workflow is auto-triggered by WhatsApp Trigger. Incoming messages are routed by the Route Types Switch: Text extracts message body directly; Audio gets media URL via WhatsApp API, downloads via HTTP Request, then transcribed to text by OpenAI audio/translate; Image gets URL via WhatsApp API, downloads via HTTP Request, then analyzed by OpenAI image/analyze (gpt-4o-mini, detail=auto) to generate description; Document gets URL via WhatsApp API, downloads via HTTP Request, then a Code node maps file extensions, and a Route Document Types Switch routes by MIME type to Extract from PDF/XLS/XLSX nodes for text extraction, ultimately formatting as a text field. All branches feed into the Knowledge Base Agent, which uses Simple Memory (sessionKey based on user's WhatsApp ID) for multi-turn context, retrieves relevant document chunks via MongoDB Vector Search (retrieve-as-tool, toolName=productDocs), generates answers with OpenAI Chat Model (gpt-4o-mini), and sends replies via the Send Response WhatsApp node.

   Setup guidance · Source text。Sticky Note1 instruction; overall workflow capability summary

   Copy prompt

   This workflow listens for WhatsApp messages (text, audio, image, documents), converts them into embeddings, searches MongoDB, and uses GPT-4o-mini to provide context-aware answers with conversation memory.

   Execution prompt · Source text。Map text prompt node (n8n-nodes-base.set); maps WhatsApp text message body to text field output to Knowledge Base Agent

   Copy prompt

   ={{ $json.messages[0].text.body }}

   Execution prompt · Source text。Map image prompt node (n8n-nodes-base.set); merges OpenAI image analyze result with image caption into text field

   Copy prompt

   =User image description: {{ $json.content }}
   
   User image caption: {{ $('Route Types').item.json.messages[0].image.caption }}

   Execution prompt · Source text。Map document prompt node (n8n-nodes-base.set); merges document extraction result with caption and MIME type into text field

   Copy prompt

   =Parsed text: {{ $json.text || $json.data || $json }}
   
   Caption text: {{ $('Route Types').item.json.messages[0].document.caption }}
   
   MimeType: {{ $('Gets WhatsApp Document Source URL').item.json.mime_type }}

   Execution prompt · Source text。Map file extensions node (n8n-nodes-base.code); normalizes MIME types for Route Document Types Switch

   Copy prompt

   let requests = $("Download Document").all()
   
   requests.forEach((request) => {
     let mime_type = request.json.mime_type
   
     if (
       mime_type === "text/calendar" || 
       mime_type === "application/ics" || 
       mime_type === "text/x-calendar"
     ) {
       request.json.mime_type = "mapped/calendar"
     }
   
     if (
       mime_type === "application/xml" || 
       mime_type === "text/xml") {
       request.json.mime_type = "mapped/xml"
     }
   
     if (!mime_type) {
       request.json.mime_type = $('Gets WhatsApp Document Source URL').first().json.mime_type
     }
   })
   
   return requests;

You are setting up a WhatsApp AI support system on n8n. Proceed in this order: 1) Confirm you have an OpenAI API key, MongoDB Atlas cluster (M10+ for vector search), Google Docs document link, and Meta WhatsApp Business account; 2) Import this template workflow into n8n; 3) Configure all credentials (Google Docs OAuth2, MongoDB, OpenAI, WhatsApp Trigger, WhatsApp API, HTTP Header Auth); 4) Create a vector search index in MongoDB Atlas (index name data_index, dimensions 1536, similarity cosine); 5) First run the document ingestion workflow manually in test environment to verify chunking and embedding; 6) Configure WhatsApp webhook to point to your n8n instance; 7) Test sending text, voice, image, and PDF messages to verify each branch processes correctly and returns context-relevant answers.