Semantic search
In one sentence
Semantic search finds results by meaning rather than by matching words, so a spoken question and the written answer can share no vocabulary at all and still be matched correctly.
Not to be confused with Knowledge graph.
Definition
Semantic search finds answers by meaning instead of by matching the exact words. A question and the answer can have no words in common and still connect.
A customer asks about walk-ins, your leaflet says no appointment needed, and the system links the two without you having to guess the phrasing.
Semantic search is the retrieval behavior; embeddings are the mechanism underneath it, and the two are best read together. Where keyword search matches the words present, with some tolerance for stemming and synonyms, semantic search compares learned representations of meaning, so paraphrases match. Keyword search excels at exact identifiers and semantic search excels at natural language questions. Neither is universally better, and production systems generally run both.
Why voice makes it more necessary
- Typed queries compress to keywords because typing is effortful.
- Spoken queries are full natural sentences because speaking is cheap.
- The vocabulary gap between how a question is spoken and how an answer is written is therefore wider in voice than in text.
- A keyword search tuned on typed query patterns will underperform badly on spoken input.
Where it fails
- Exact identifiers. Product codes, postcodes, part numbers and dates. Near in meaning is not near in string.
- Negation. The phrase we do not offer parking can land uncomfortably close in meaning space to we offer parking.
- Rare specific terms with little training representation.
- Highly specific numeric queries.
Hybrid search is the production answer
- Run semantic and keyword search in parallel and merge the results.
- The merge, typically a weighted fusion of the two ranked lists, is itself a tuning decision.
- This handles both natural language questions and exact identifiers, which is what a business corpus actually contains.
- Semantic search is the retrieval step in a RAG system: it supplies candidates, reranking orders them, and the model composes from them. Retrieval quality bounds answer quality, which makes it the highest-leverage component after chunking.
Common misconception
That semantic search understands the question. It compares learned representations of meaning. That works impressively well and fails in characteristic ways, particularly around negation and exact strings, and calling it understanding hides exactly those weaknesses.
Why it matters commercially
Semantic search is what makes it unnecessary to anticipate the words customers will use. A business writes its documents once, and visitors find the answers regardless of phrasing. Without it, every way a question might be worded has to be foreseen and written out, which no real corpus ever manages.
In voice specifically
The gap it closes is wider in speech. A typed query is a keyword; a spoken query is a sentence. The same corpus that a keyword search covers adequately for typed traffic will strand a large share of spoken questions, because the caller says it in full sentences the documents never used.
Where AsqVox fits
This is the mechanism by which a spoken question reaches the right passage of the documents a business has uploaded, without the business having to predict the phrasing in advance.
Visual
Finding the answer that shares no words
One path matches strings and finds nothing. The other matches meaning and finds the passage.
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.
Semantic search compares vector representations of meaning, typically using cosine similarity, which is scale-invariant and therefore insensitive to passage length.
cosine similarityindustry rangeStandard vector retrieval convention, 2026 - Scale invariance is why a short heading and a long paragraph can both match a question fairly, rather than the longer passage winning on raw magnitude.
Hybrid retrieval, combining semantic and keyword search, is standard production practice specifically because embeddings handle exact identifiers poorly.
semantic plus keywordindustry rangeProduction RAG engineering convention, 2026 - Read a spec sheet that offers only one of the two as covering half the queries a business corpus actually receives.
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. A misheard question is embedded faithfully and retrieves the wrong passage.
2.7% against 8 to 12%independentWhisper Large-v3 published benchmark and reported real-world results, 2026 - Speech recognition error compounds retrieval error. Semantic search will confidently match whatever it was handed, correct transcription or not.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate in production conditions.
5.26 to 6.84%vendor claimDeepgram, Nova-3 reported figures, 2026 - Vendor-reported, so a supplier best case. It still sits above the clean benchmark, which is the input semantic search has to work from.
Retrieval sits inside the sub-second time to first audio budget, and above roughly 1,500 milliseconds a conversation feels broken.
within ~1,500 msindustry rangeVoice interaction latency convention, 2026 - Semantic search cannot be tuned for quality in isolation. Every extra millisecond of lookup is heard as silence before the agent speaks.
There is no reliable transferable benchmark for semantic search quality. Retrieval performance depends on corpus, query distribution and chunking, and public leaderboard position correlates only loosely with performance on a specific business corpus.
-no reliable figureTreat a vendor leaderboard rank as directional at best. The only number that matters is retrieval quality measured on your own documents and your own callers.
Examples
In practice
A clinic describes its evening service as requiring no appointment. Visitors ask about walk-ins. Keyword search returns nothing because the documents never use that word. Semantic search matches correctly on meaning. The clinic did not have to anticipate the phrasing, which is the entire practical value.
The everyday version
Semantic search means the system matches on meaning rather than on exact words. Your customer asks about walk-ins, your leaflet says no appointment needed, and it connects the two. You do not have to guess every way someone might phrase a question.
Usage
Who says it
- Search and AI engineers use it precisely.
- It appears in buyer-facing material more often than embeddings does, and is the better term for a non-technical audience.
Where it turns up
- Next to retrieval method, hybrid search support, supported languages and search configuration on a spec sheet.
- A serious RFP asks how exact identifiers are handled, because that is the query class pure semantic search misses.
Common misuse
- Presenting it as a complete replacement for keyword search.
- Assuming it handles exact identifiers well.
- Describing it as understanding, which obscures the negation weakness.
Questions people ask
What is the difference between semantic search and keyword search?
Keyword search matches the words present, with some tolerance for stemming and synonyms. Semantic search compares learned representations of meaning, so a paraphrase matches even when it shares no words with the source. Keyword search wins on exact identifiers and semantic search wins on natural language questions, which is why production systems run both.
Where does semantic search fail?
On exact identifiers such as product codes and postcodes, where near in meaning is not near in string; on negation, where the phrase we do not offer parking sits close in meaning to we offer parking; on rare specific terms with little training representation; and on highly specific numeric queries. Hybrid search exists to cover these gaps.
Why does voice need semantic search more than text does?
Typing is effortful, so typed queries compress to keywords. Speaking is cheap, so spoken queries are full sentences. The gap between how a question is spoken and how an answer is written is wider in voice, and a keyword search tuned on typed patterns underperforms badly on speech.
Does semantic search understand the question?
No. It compares learned representations of meaning. That works impressively well and fails in characteristic ways, particularly around negation and exact strings. Calling it understanding hides exactly those weaknesses, which matters when a wrong match sounds as confident as a right one.
Last reviewed 3 August 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.