Embeddings

Embeddings are mathematical representations of text in the form of vectors of numbers. In simple words, they convert words, phrases, or entire messages into lists of numbers that capture their meaning. Thanks to this, an AI system can compare texts and understand how similar they are, even if they use different words to say the same thing.

Why do they matter for your SME?

If you ever wanted a WhatsApp bot to understand that 'how much does shipping cost?' and 'do you charge for delivery?' are the same question, you need embeddings. Without them, traditional chatbots rely on exact keywords and break down with any variation. With embeddings, the AI can group messages by topic, search for answers similar to ones you have already given, and suggest the best response to you even if the customer writes however they want.

Concrete example

Imagine you have a clothing store and a customer writes to you: 'hi, do you have the black t-shirt in size M?'. With embeddings, the system understands that this is a product availability inquiry, not a complaint or a claim. Furthermore, if you have already responded before 'yes, we have it, we will hold it for you' to a similar message, the AI suggests it to you as a response, even if the customer wrote 'is there any medium black t-shirt left?'.

Common mistakes

• Thinking embeddings are the same as keywords: no, they capture context, not just terms.
• Believing that a chatbot with embeddings responds on its own without supervision: the AI suggests, but a human approves.
• Expecting them to work without training: they need real business data to learn.

How it's used in practice

In a real business, embeddings work behind the scenes. You don't see them, but you feel them in your message inbox.

Example: a real estate agency receives 'Is the 2-bedroom apartment in Palermo still available?' and 'Do you have anything with 2 bedrooms in Palermo?'. Without embeddings, they are two different messages. With embeddings, the system groups them as the same intent and suggests a response based on what you've already answered before.

The AI doesn't respond on its own: it suggests a response and you approve it with a click. If the customer writes 'and the price?' or 'does it include HOA fees?', the system understands that the same conversation continues and adjusts the suggestion.

The important thing: embeddings are not magic. They learn from your real responses. The more you respond, the better the suggestions become.

• You receive a new message in the inbox.
• The AI compares it with previous messages using embeddings.
• It suggests a response that worked in a similar case.
• You approve it, edit it, or discard it.

What it's confused with and how it differs

Embeddings are confused with two things: flow chatbots and language models.

A flow chatbot works with rules: if the customer says 'hello', reply 'hello'. If they say 'price', reply 'the price is X'. It is programmed by hand and breaks with any variation. Embeddings don't require you to program every path: they understand the meaning of the message.

They are not a language model either. The model generates text; embeddings are the numerical representation that allows comparing meanings. In Wando, the model uses embeddings to understand and suggest, but always with a human who approves.

The key difference: a flow is a tree that is built and maintained. Embeddings are a layer that learns on its own from your responses, without you having to touch anything.

What happens if you ignore it

If you don't use embeddings, your customer service depends on the customer typing exactly the words you expect.

Result: automatic responses that don't understand what the customer wants, customers who repeat the question three times and end up leaving for the competition.

With a flow chatbot, every variation of a question is a case you have to program. If a customer writes 'hello, can you give me the shipping price?' and your flow only recognizes 'price', it won't handle them well.

Embeddings make variations cost you no effort. Without them, your team ends up answering the same thing twenty times a day, and the bot learns nothing from those responses.

The cost of ignoring it is not technical: it's your team's time and customers who leave without buying.

How it relates to Meta's official API

Embeddings are not part of Meta's WhatsApp API. They are an AI technology that works separately.

Wando uses Meta's official API (WhatsApp Cloud API) to connect to your WhatsApp Business. Embeddings operate on Wando's side: they take incoming messages, convert them into vectors, and search for response suggestions.

Meta's API handles sending and receiving messages, templates, and conversation pricing. Embeddings handle text understanding.

In practice, you don't need to configure any of this. Wando combines Meta's API with embeddings to make the inbox work. Meta's conversation prices vary by country; you can check them in the official documentation.

The important thing: embeddings don't replace Meta's API, they complement it.

When it does NOT apply or is not advisable

Embeddings are not the solution for everything. If your business receives ten messages a day and you answer them all right away, maybe you don't need AI for that.

They are also not useful if you don't have a history of responses to learn from. AI improves with real data; without it, suggestions are generic.

If your customer service is 100% human and you don't want a system to suggest replies, embeddings bring you nothing.

And beware: embeddings don't understand the physical world. If a customer asks about a product that isn't in your catalog, the system won't make it up. You need someone to add it.

In those cases, a simple chatbot or just answering manually can be more effective and cheaper.

Common mistakes of those who just understand it

Thinking that embeddings are a database of synonyms. It's not just 't-shirt' = 'tee': it's understanding that 'do you have the black one?' refers to the shirt they were talking about.

Believing that with embeddings the bot answers on its own. In Wando, AI suggests and a human approves. Without that oversight, errors multiply.

Expecting it to work perfectly from day one. Embeddings need real examples. The more approved responses, the better the suggestion.

Confusing embeddings with the Meta API. They are different things: one is the connection with WhatsApp, the other is the understanding of text.

The most expensive mistake: not testing. A 14-day trial without a card tells you whether your business benefits or not.