An embedding is a list of numbers that represents the meaning of a text. Two texts that say similar things get similar lists, even if they share no words. That is what makes semantic search possible: someone who types “I cannot get into my account” finds the article “recovering access”, which does not contain the word “get into”. It is the piece underneath assistants that answer from documents (RAG), but it also works on its own.
Searching by words or by meaning
|
By words (the normal kind) |
By meaning (embeddings) |
| How it finds |
Looks for the written words |
Compares the meaning of the question with that of the texts |
| Good for |
Codes, references, exact names, invoice numbers |
Natural-language questions, synonyms, vague descriptions |
| Fails on |
People who use other words for the same thing |
Exact codes and names: it may return “similar” things instead of the one asked for |
| What it needs |
Nothing special (WordPress search, the database) |
An embedding model and somewhere to store the numbers |
| Cost and privacy |
Stays on your server |
The text goes to a model, a provider’s or your own |
The best result usually comes from combining the two: word search for what is exact and meaning search for the rest.
What it is for on a small site
| 1 |
Site search that understands questions. The visitor types the way they speak and finds the right page.
|
|
| 2 |
“Similar articles” at the end of each article, without choosing them by hand.
|
|
| 3 |
Routing support requests: compare a message with the frequently asked questions and suggest the closest answer.
|
|
| 4 |
Finding duplicates or repeated texts on a large site.
|
|
What you need to build
| 1 |
Cut the content into pieces that each deal with one thing (a section, a question).
|
|
| 2 |
Generate one embedding per piece, with a model made for that. Answering models and embedding models are different; each has its own pricing page.
|
|
| 4 |
At search time, turn the question into an embedding with the same model and fetch the closest ones.
|
|
| 5 |
Redo the embeddings when the text changes. New text with old numbers gives wrong results.
|
|
Quality depends a lot on how you cut the text. Pieces that are too small lose context; pieces that are too big mix subjects. Try real questions and look at what the search returns before connecting it to the site.
|
Mixing models ruins everything. One model’s numbers cannot be compared with another’s. If you change model, you have to redo all the embeddings.
|
|
For a site of a few dozen pages, it may not be worth building anything: a good frequently asked questions page and a well-tuned normal search solve most of it. See RAG explained.
|
|
Want to build a search like this on a server of your own? Have a look at the VPS plans.
See VPS plans
|
RECOMMENDED PRODUCT Web hosting with cPanel Domain and SSL included, daily backups and the panel you already know. from $6.59/mo (3-year plan, with coupon) See plans |
Português
English
