Embeddings and meaning-based search in plain English

An embedding is a list of numbers that stands for the meaning of a piece of text. Search tools compare those lists, not spellings. So a question about "time off" can find a page that says "annual leave." Odd, but it works.

The problem it solves

Old-style keyword search looks for the words you typed. Microsoft's documentation says traditional keyword search fails when queries don't match document terminology.¹ It gives a sample where users ask about "PTO" but the documents say "time off," "telecommute" and "recent hires."¹

People rarely phrase a question the way a document was written. Meaning-based search, also called semantic or vector search, is built for that gap.

What an embedding is

Wikipedia describes a word embedding as a representation, typically a real-valued vector, that encodes the meaning of the word.² A vector is just an ordered list of numbers. Words that are closer together in this number space are expected to be similar in meaning.²

Microsoft gives a plain example: "dog" and "canine" are conceptually similar but linguistically distinct.³ Different spelling, close meaning.

Assistants usually apply this idea to whole chunks of text, not single words. How documents are split explains the chunks.

The steps behind a lookup

• Each chunk of your documents is converted into its number list and stored.
• When you ask a question, the question is converted the same way.
• The system looks for stored lists nearest to the question's list. Microsoft says its search uses a nearest neighbors method to place similar vectors close together in an index.³
• The text of the closest chunks is handed to the model.

Microsoft puts the outcome simply: vector similarity search matches concepts, not exact terms.¹ Matching also works across languages and other content types, according to Microsoft's overview, for example "dog" in English and "hund" in German.³

Where it can miss

Meaning-based search isn't magic. Anthropic says embedding models excel at capturing semantic relationships but can miss exact matches.⁴ Its example is a support database where someone searches for an error code. The model might find content about error codes in general. It could still miss the exact code.⁴

For a small business, think of part numbers, SKU codes, invoice numbers, staff initials and unusual product names. Anthropic says an older exact-wording technique is particularly effective for queries that include unique identifiers or technical terms.⁴

Many tools therefore run both kinds of search at once. Microsoft calls this hybrid search and says it often provides better results than vector or keyword search alone.³ Whether your assistant does this is a question for its documentation.

What this means for your files

• Plain wording helps. Say what a thing is in the same words a customer would use, and add the formal name too.
• Put codes and exact names in the text, so exact-match search has something to hold.
• If a question gets nothing, try asking it a second way. If one wording works and the other doesn't, matching is the likely culprit.

A good match doesn't guarantee a good reply, and why answers go wrong looks at what can happen after the lookup.