Retrieval-augmented generation (RAG) is the technique that allows an AI agent to base its responses on your specific knowledge base—PDFs, website, catalogs—instead of responding with its general, trained knowledge. For a WhatsApp chatbot in 2026, RAG is the difference between an agent that recommends your actual products and policies versus one that conjures up competitor names and made-up prices.
What is RAG in simple terms?
When a customer writes to your chatbot "What is the return policy?", there are two ways to respond:
- Without RAG: The LLM responds based on their trained general knowledge — something generic like "typically return policies are 30 days."
- With RAG: The system first searches in YOUR return policy document, retrieves the exact ticket and responds "According to our policy, you have 14 calendar days from delivery to return the product unused, with tags and original packaging."
RAG brings the precise and specific AI agent to your business.
How RAG works step by step
- Intake. You upload your documents (PDF, DOCX, website). The system breaks them down into passages of 200-500 words.
- Embedding. Each passage becomes a numerical vector (an "embedding") that captures the semantic meaning.
- Storage. The vectors are stored in a vector database (MongoDB Atlas, Pinecone, Azure AI Search).
- Consultation. When a question arrives, it also becomes a vector and the closest semantically relevant passage is searched for.
- Generation. The relevant passages are sent to the LLM along with the question. The LLM responds based on those passages.
All of this happens in less than 1 second per response.
Types of searches in RAG
| Guy | How it works | When to use |
|---|---|---|
| Vectorial (semantics) | Search by meaning | Vague or rephrased questions |
| BM25 (keywords) | Search by exact terms | Questions with specific terms |
| Hybrid (vectorial + BM25) | Combine both with reciprocal rank fusion | The standard in 2026 |
Aurora Inbox uses hybrid vector search + BM25 for maximum accuracy.
What documents to upload to RAG
Minimum recommended for a WhatsApp chatbot:
- FAQ complete of your company.
- Pricing page and plans.
- Returns policy / warranty / shipping.
- Product documentation or main services.
- Terms and conditions legal.
- Procedures Manual internal (if applicable).
- Product catalog with detailed descriptions.
Aurora Inbox supports up to 40 PDFs / DOCX / XLSX either 40 URLs per site crawled per agent. More than 40 native languages.
RAG vs fine-tuning: which is better in 2026?
| Appearance | RAG | Fine-tuning |
|---|---|---|
| Knowledge update | Replace document, done | Retrain model, days |
| Cost | Under | High |
| Implementation time | Hours | Weeks |
| Traceability of responses | Yes (quote passage) | No |
| Typical use cases | Business information | Very specific tone or style |
For WhatsApp chatbot case 95%, RAG is the correct answer. Fine-tuning is only for very specific cases where you need a specific tone or vocabulary.
Common mistakes when implementing RAG
- Upload only prices without descriptions. The RAG does not understand context without sufficient text.
- Outdated documents. If the policy has changed and you don't update it, the agent will quote the old one.
- Mix languages in a single document. It's best to have one document per language for a clean embed.
- Very short passages (<100 words). Embedding loses context.
- Very long passages (>500 words). The LLM does not focus well.
- Don't try with real questions. What goes up to the RAG is not always recovered with the typical query.
How Aurora Inbox implements RAG
Aurora Inbox has RAG embedded for each AI agent:
- Upload documents from the UI — PDFs, DOCX, XLSX, or site URLs.
- Automatic processing — 300-word chunks, embeddings via Azure OpenAI, vector storage in MongoDB Atlas.
- Hybrid search vector + BM25 with reciprocal rank fusion.
- Traceability — the agent may cite the source document.
- More than 40 languages natively supported.
- Automatic re-indexing when you replace a document.
Without programming, without its own infrastructure.
Why Aurora Inbox
Aurora Inbox combines native onboard RAG + real LLM agent (GPT-5) + a navigable catalog + scheduling + multichannel support in a single platform. You upload your documents and the agent responds based on them in less than a day.
Frequently Asked Questions
Why do I need a RAG if I have an agent with an LLM?
Without RAG, the LLM responds with their general knowledge—they might speculate or ignore specific details about your company. RAG bases its answers on your actual documents.
How many documents can I upload to RAG?
Aurora Inbox supports up to 40 files per agent. Beyond that, you can create multiple specialized agents.
Does RAG support languages other than English?
Yes. Aurora Inbox supports over 40 native languages via GPT-5.
What happens if I update a document in RAG?
Aurora Inbox automatically re-indexes it in minutes. The agent starts using the new version without downtime.
Is RAG more expensive than fine-tuning?
On the contrary. RAG is cheaper, faster to implement, and easier to maintain.
Does RAG work with large product catalogs?
Yes. Aurora Inbox supports catalogs of up to 9,000 products, navigable by AI with semantic search.
