Embeddings
In one sentence
An embedding is a list of numbers representing the meaning of a piece of text, arranged so that two passages about the same thing sit close together even when they share no words, which is what lets a system search by meaning instead of by keyword.
Not to be confused with Vector database.
Definition
An embedding is a run of numbers standing in for what a piece of text means. Write about the same thing twice, in completely different words, and the two runs of numbers still come out looking alike.
That is what makes searching by meaning possible, rather than searching for the exact words someone happened to type.
What an embedding model does is place text inside a vector space of high dimension, typically somewhere between several hundred and a few thousand axes, arranged so that distance stands in for difference in meaning. The usual way to read that distance is cosine similarity, which measures the angle between two vectors and ignores their magnitude, so a short passage and a long one can be compared on equal terms.
Why this matters more in voice than in text
- Someone asks whether you take walk-ins on a Thursday evening. What the document actually says is that no appointment is necessary for the late clinic. Not one word in common. A keyword index has nothing to offer. An embedding index hands back the right passage.
- Spoken questions run longer, sound more conversational and vary more in phrasing than typed ones. That widens the gap between how the question is worded and how the answer was written, which makes semantic matching more valuable in voice than in text search rather than less.
The properties that decide practical quality
- Dimensionality. More dimensions can hold more nuance and cost more to store and search. It is a genuine engineering tradeoff, not an axis where bigger wins.
- Domain fit. A general-purpose model may fail to separate terms of art in a specialized field, where the distinctions that matter are precisely the ones it was never trained to notice.
- Asymmetry. Questions and documents are not the same genre of writing. Certain models train a dedicated encoder for each, and doing so lifts retrieval noticeably above running both through one.
- Multilingual behavior. A model may or may not put the same sentence in two languages in the same neighborhood, and only the ones that do can retrieve across languages. In any market where the question comes in one language and the source material sits in another, the difference is enormous.
What embeddings cannot do
- They handle exact identity poorly. Product codes, part numbers, postcodes and dates are badly served by semantic similarity, because near in meaning is not the same as identical as a string. Hybrid search, running embedding search alongside keyword search, exists for that reason and is usually the right production answer.
- They do not reason. Proximity is not inference.
Common misconception
That embeddings understand text. What they hold is statistical regularity in how meaning gets expressed, absorbed from training data. It performs remarkably well, and it breaks in predictable places, negation being the one you will meet first: "we do not offer parking" and "we offer parking" can end up sitting uncomfortably close.
Why it matters commercially
Embedding quality sets the ceiling on retrieval quality, and retrieval quality sets the ceiling on answer quality. It is the least visible choice in a RAG system and one of the most consequential, and in most deployments it was never really made: the default was accepted and nobody went back to it.
In voice specifically
A visitor typing into a search box edits themselves. A visitor speaking does not. Questions arrive with false starts, filler and whatever phrasing came to mind, and none of it will match the wording of the document. Semantic matching absorbs that variation, which is why voice leans on embeddings harder than a search box ever has to.
Where AsqVox fits
True RAG over uploaded documents, behind the Orb, runs on embeddings. They are what lets a visitor phrase a question however it comes out and still land on the right passage of a business's own material.
Visual
Finding the answer that shares no words with the question
Meaning, converted into arithmetic.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
No single authoritative accuracy figure exists for embedding models. Retrieval quality depends on the corpus, the distribution of queries, the chunking strategy and the domain.
-no reliable figureA vendor quoting one number for embedding quality is quoting a benchmark run on data that is not yours. The only figure worth acting on is one measured against your own corpus and your own questions.
Public retrieval benchmarks exist and are useful for coarse comparison, but leaderboard position correlates only loosely with performance on a specific business corpus.
-no reliable figureWorth saying plainly rather than implying. A leaderboard narrows the shortlist. It does not settle the choice.
Cosine similarity is the standard proximity measure for text embeddings, because it is scale-invariant and therefore insensitive to how long the passage is.
Cosine similarityindustry rangeStandard practice in text retrieval, 2026 - An established convention rather than a measured finding. It is why a two-line answer and a two-page policy can be compared against the same question without the longer one being penalized.
Hybrid retrieval, combining semantic search with keyword search, is standard production practice, specifically because embeddings handle exact identifiers poorly.
Semantic plus keywordindustry rangeEngineering consensus, 2026 - Consensus rather than a measured finding. Present it that way. The practical trigger is any corpus containing part numbers, SKUs, postcodes, policy numbers or dates.
Whisper Large-v3 achieves around 2.7 percent word error rate on clean benchmark audio and roughly 8 to 12 percent on real-world English.
2.7% clean, 8 to 12% real-worldindependentPublished benchmark and field-reported ranges, 2026 - Relevant here because embeddings sit downstream of speech. A misheard question is embedded faithfully and retrieves the wrong passage, and no amount of embedding quality recovers from a transcription error upstream.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate in production.
5.26 to 6.84%vendor claimDeepgram, 2026 - Vendor-reported, so read it as a vendor figure. The point of quoting it beside the Whisper range is the spread between clean benchmark conditions and production, which is the error the retrieval step inherits.
Examples
In practice
A team finds retrieval failing on queries that contain part numbers, while descriptive questions perform well. The embedding model treats "AX7-99Q" and "AX7-99C" as near-identical, because character-level distinctions carry almost no semantic weight. Adding a keyword search path and merging its results with the semantic path resolves it. The embedding model was not the wrong choice. It was the wrong tool for that one query type.
The everyday version
Embeddings are how the system works out that "are you open late on a Thursday" and "when does the evening clinic finish" are the same question. You do not have to write your documents in the exact words your customers will use, because the match happens on meaning rather than on wording.
Usage
Who says it
- ML and search engineers, constantly and precisely.
- Increasingly in buyer-facing material, usually under the name semantic search, which is the better term for a non-technical audience.
- Procurement meets it only through data-handling questions: where the vectors are stored, and whether the source text can be reconstructed from them.
Where it turns up
- Usually in the same clause as retrieval method, hybrid search support, supported languages, dimensionality, and whether embeddings are generated on the vendor infrastructure or a third party.
- The third-party question is the one that matters for data residency. Generating embeddings normally means sending the source text to the embedding provider.
Common misuse
- Treating embedding search as a complete replacement for keyword search. Exact identifiers still need keyword matching.
- Assuming multilingual capability. Cross-lingual retrieval is a property of specific models, not of embeddings in general.
- Describing embeddings as understanding. They encode similarity, and negation is where that framing visibly breaks.
Questions people ask
What is an embedding in AI?
It is a list of numbers standing for the meaning of a passage of text. An embedding model maps text into a space of several hundred to a few thousand dimensions, where two passages about the same thing land near each other even if they share no words. Closeness is normally measured with cosine similarity, the angle between two vectors, which is insensitive to how long the passage is.
What is the difference between semantic search and keyword search?
Keyword search matches strings. Semantic search matches meaning, so a question about walk-ins on a Thursday evening can retrieve a passage saying no appointment is necessary for the late clinic, with no shared words. Keyword search still wins on exact identifiers such as part numbers and postcodes, which is why hybrid search running both together is standard production practice.
Do embeddings understand text?
No. They encode statistical regularities of meaning learned from training data. It works impressively well and it fails in characteristic ways, most visibly around negation, where "we do not offer parking" and "we offer parking" can land uncomfortably close together. Proximity is not inference.
Do embeddings work across languages?
Only when the model was trained for it. Some models place equivalent text in different languages near each other and support cross-lingual retrieval, and others do not. It is a property of the specific model rather than of embeddings in general, and assuming it is a common and expensive mistake in markets where questions arrive in one language and documents exist in another.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.