Embeddings, an embedding model, and a vector database are enough for a demo. They are not enough for a search product.

A user searches for “how to handle database connection timeouts”, and the top result is a section about HTTP request timeouts. It is related, but it is not what they were looking for. Another user searches for “Python list comprehension”, and the best match is a paragraph about Python decorators that happens to mention lists.

In both cases, the system understands the general topic, but misses the user’s actual intent.

The gap between "returns something related" and "returns the right evidence" is where most retrieval engineering happens. Closing it requires careful chunking, hybrid retrieval, reranking, metadata design, and measurement.

That is what this chapter is about.