Hébergement VPS n8n | Automatisation sécurisée et évolutive
Formation n8n en Français : Automatisation & Agents IA
This course rebuilds the classic "WhatsApp Sales Agent + RAG" template with one major upgrade: the agent now understands voice notes, photos and videos, not just text. Whatever a customer sends, it becomes plain text before reaching the AI Sales Agent, which can also search your product catalogue (RAG) and remembers the last messages per customer.
PART 1 - Build the knowledge base (run once)
Manual Trigger -> Knowledge Base Settings -> Download Product Brochure -> Extract Brochure Text -> Insert Product Catalogue into Vector Store
(subnodes: Embeddings for Indexing, Load Brochure Document <- Split Brochure Text)
PART 2 - WhatsApp Sales Agent (production)
WhatsApp Trigger -> Configuration -> Route by Message Type (Switch)
case Audio -> Get Audio Media URL -> Download Audio File -> Transcribe Voice Note -----\
case Image -> Get Image Media URL -> Download Image File -> Analyze Image with GPT Vision -\
case Video -> Get Video Media URL -> Download Video File -> Describe Video with Gemini -----> Prepare Agent Input -> AI Sales Agent -> Reply to Customer
case Text ------------------------------------------------------------------------------/
fallback (Unsupported) -> Reply Unsupported Message
AI Sales Agent subnodes: OpenAI Chat Model, Per-Customer Chat Memory, query_product_catalogue (tool) <- Product Catalogue (Retrieve) <- Embeddings for Retrieval
| Credential | Type | Where to get it | Used on |
|---|---|---|---|
| WhatsApp Trigger credential | whatsAppTriggerApi |
Meta developer app > WhatsApp product > API setup (Client ID/Secret) | WhatsApp Trigger |
| WhatsApp Business credential | whatsAppApi |
Same Meta app, permanent access token | All WhatsApp send / media nodes, and the 3 HTTP download nodes (predefined credential type) |
| OpenAI account | openAiApi |
platform.openai.com > API keys | Transcription, Vision, Chat model, Embeddings |
| Google Gemini (PaLM) API key | googlePalmApi |
Google AI Studio > Get API key | Video understanding |
Test the WhatsApp credential: send yourself a text message on your test number after activating the workflow; you should see an execution appear.
Node type: Set (n8n-nodes-base.set v3.4), includeOtherFields: true (critical - keeps the original WhatsApp payload flowing through).
| Name | Type | Default value | Role |
|---|---|---|---|
| openai_model | string | gpt-5.4-mini |
Chat model used by the AI Sales Agent |
| vision_model | string | gpt-5.4-mini |
Model used to describe photos |
| video_model | string | models/gemini-3.5-flash |
Gemini model used to describe videos |
| memory_context_window | number | 10 |
How many past messages the agent remembers per customer |
| max_reply_length | number | 1500 |
Hard cap on WhatsApp reply length |
| vector_store_key | string | product_catalogue |
Must match the key used in "Knowledge Base Settings" |
| phone_number_id | string | (you fill this) | Your WhatsApp Business phone number ID |
| system_prompt | string | (provided, editable) | The Sales Agent persona and instructions |
| product_catalogue_description | string | (provided) | Description shown to the LLM for the RAG tool |
| unsupported_message_text | string | (provided) | Reply sent for message types we don't handle |