Comment by Ahmet Özel on "Vector Databases & Embeddings 101"

Ahmet Özel

AI Engineer. Computer Vision, RAG and LLM agents.

The money-back versus refund example is the cleanest way to motivate this, and it is worth keeping the counter-example beside it so readers do not over-correct.

Semantic search fixes the vocabulary mismatch and introduces a new one: a user searching for error code ERR-1042 or an exact part number gets documents about errors in general, because a short alphanumeric token carries almost no semantic signal. Keyword search never had that problem precisely because it never tried to understand anything.

Most production setups end up running both and fusing the results, so framing embeddings as the thing that covers what lexical search misses, rather than as its replacement, sets people up better.

Nagappan S

Building scalable products at the intersection of Full Stack Engineering & AI

Sep 21

Hey Ahmet Özel Thank you so much for taking your time to read through & drop your comments

One practical detail for that hybrid setup is preserving identifiers before normalization. Lowercasing ordinary prose may be harmless, but stripping punctuation or converting a zero-prefixed code into a number can change the identifier the lexical branch is meant to protect.

I would keep the original value alongside normalized text and add a few identifier round-trip checks to ingestion. Then the vocabulary-mismatch example and the exact-match example can remain separate tests, with neither branch quietly losing the information it was introduced to handle.

Search Hashnode