Embeddings

An embedding is a representation of a piece of data (a word, a phrase, a document, an image) converted into a vector, that is, a list of numbers. The key is that this vector captures meaning: texts about the same thing end up with vectors close to each other, even if they use different words. So "unpaid invoice" and "outstanding bill" end up close, while "unpaid invoice" and "vacation" end up far apart.

That closeness is measured with geometric distance (most commonly cosine similarity). Thanks to that, a machine can compare the meaning of two things without understanding the language the way a person does: it just compares numbers. It is the piece that makes semantic search possible (searching by idea, not by exact word matches) and the first step of patterns like RAG.

In data analytics, embeddings let you connect free text (tickets, reviews, product descriptions) with natural language queries, a capability that is part of what Metrix enables when it works with unstructured data.

Why embeddings matter

Computers do not understand words; they understand numbers. For decades, searching for information meant exact text matching: if you searched for "cell phone" you would not find "smartphone", even though they are the same thing. Embeddings break that limitation because they translate meaning into coordinates. A model trained on huge volumes of text learns to place each concept in a space with many dimensions (hundreds or thousands), so that the distance between two vectors reflects how similar their meanings are.

The result is that you can ask in your own words and retrieve what is relevant even when the wording does not match. That is the basis of semantic search, recommendation systems, duplicate detection and, above all, of how generative AI finds the right context before answering.

How it works, step by step

1. Input: you take a text (or an image, audio, etc.).
2. Embedding model: a specialized model turns it into a fixed-size vector, for example 1,536 numbers.
3. Storage: that vector is stored in a vector database, along with thousands or millions of others.
4. Query: when a question comes in, it is also converted into a vector.
5. Nearest-neighbor search: the system compares the query vector against the stored ones and returns the closest ones (the most similar in meaning).

That "find the closest" mechanism is what powers the retrieval stage in RAG, where the model first finds the relevant documents and only then writes the answer based on them.

A concrete example (Argentina, B2B)

Picture a consumer goods distributor in Buenos Aires with 40,000 accumulated customer service tickets. With traditional search, finding "late delivery problems" requires guessing the exact words each customer used. With embeddings, the Customer Success team vectorizes all the tickets once; after that, a query like "customers upset because the order arrived late" retrieves cases that said "I still haven't received anything", "the delivery truck never showed up" or "dispatch delay", even though none of them uses the word "late". The same applies to a catalog: searching for "running shoes for the rain" can return products described as "waterproof athletic footwear", raising conversion without touching the text of each product page.

Common mistakes

• Confusing an embedding with an LLM: the embedding model only produces vectors; it does not generate text. It is a different component from the LLM that writes the answer.
• Mixing models: comparing vectors created with different models makes no sense; the coordinates are not compatible. You need to keep a single embedding model per collection.
• Forgetting to re-index: if you switch models, you have to regenerate all the embeddings, not just the new ones.
• Expecting perfect precision: semantic similarity is approximate; it is best combined with filters and business rules.

Embeddings vs keyword search

In practice, the best systems use a hybrid approach: they combine lexical search (exact, fast, ideal for product codes or invoice numbers) with semantic search (flexible, ideal for natural language), and so they get the best of both worlds.

Where it shows up day to day

Embeddings are behind features you already use without noticing: an e-commerce site's internal search engine, the deduplication of similar leads in a CRM, the automatic classification of documents by topic, the chatbots that answer from the company's knowledge base and the dashboards that let you ask questions in natural language about qualitative data. They are a quiet but decisive piece of infrastructure: without them, almost none of the AI applied to text would work with the naturalness we have come to expect.