RAG for WhatsApp: a chatbot that answers with updated documents

Published on 2026-11-16 by ArduBot

How to use RAG with WhatsApp Business to query policies and catalogs, cite sources, and transfer sensitive inquiries to people.

Your team changes a shipping condition, but the WhatsApp bot keeps giving the previous answer. This happens when information is copied into static prompts or when a flow queries documents without version control. RAG allows searching for the relevant information upon receiving each question.

The journey of a question

Imagine a company that sells spare parts. A customer asks: “Does this component have a warranty if it was installed by a third party?”. The WhatsApp channel receives the message, identifies the product and allowed context, queries a warranty document database, and retrieves the corresponding section. A model can explain it in plain language and cite the published policy.

The conceptual flow would be: WhatsApp Business webhook -> intent identification -> RAG search -> evidence validation -> response or transfer to an operator. Retrieval does not need to store the customer's entire history in the index.

Differentiating knowledge from operations

RAG works well with manuals, FAQs, warranty conditions, and instructions. It is not the ideal source for knowing if an order has already been shipped, how much stock is left now, or what balance a person has. Those queries require authenticated APIs from the order system or the CRM.

A solid design uses the retriever for documentary knowledge and separate tools for live data. When the question mixes both (“can I return my order 123?”), the policy is queried via RAG and the identity/order status is checked in the transactional system.

Three measures to avoid errors

Documents with validity: add effective date, language, and product. When a new policy is published, the previous one is deactivated or marked as historical.

No answer is also an answer: if the index does not find a clear clause, the bot must not promise an exception. It can ask for a piece of information or transfer the case to a person.

Privacy: do not allow an unauthenticated search to access internal documents. Before exposing an invoice or personal data, you must verify identity and authorization.

An example of a verifiable response

Question: “How long do I have to return it?”. Expected response: “According to the current policy for online purchases, the deadline is the one indicated in the ‘Returns’ section [source]. To confirm if it applies to your order, I need to verify the purchase.” We do not invent a number of days here: the actual duration must come from the official document.

Also measure the proportion of responses that end in human intervention, how many citations are correct, and how many customers repeat their question because the bot did not understand it.

How to launch a small test

Start with a single product family and 20-30 frequently asked questions. Use a controlled document collection, record the sources, and review errors weekly. If the test only works for prepared questions, it still does not prove production quality.

Business experiment: ArduMaker is studying an automatic RAG integration to retrieve documents via API. If you maintain a WhatsApp chatbot and are interested in connecting catalogs or manuals without retraining for every change, you can explain your case from ArduMaker.

To read more: Amazon Bedrock RAG · WhatsApp Business API for businesses.

  • RAG
  • WhatsApp
  • chatbot
  • atención al cliente
  • base de conocimiento

Keep reading

All articles · VPS plans