Home / Product / Semantic search
Semantic search
Three retrieval modes — keyword, semantic and hybrid — over the same index and the same filters. All three are available on every plan, including the free one.
Why exact matching fails on news in particular
Keyword search asks your user to guess the words a newsroom chose. In most domains that is a mild inconvenience. In news it is a systematic failure, because the vocabulary is set by whoever wrote the headline and it changes between editions of the same story.
A layoff is a layoff, a restructuring, a headcount reduction, a workforce realignment or letting people go, depending on whether the source is the company, a trade publication or a union statement. A search for one of those terms returns a slice of the coverage that correlates with editorial stance rather than with events, which is worse than returning nothing: it looks like a complete answer.
The same problem appears with proper nouns and gets sharper outside English. Indian publications transliterate names three ways in a single news cycle. Japanese renders company names in kanji, in katakana, in romaji or as a ticker, sometimes within one article. Arabic is published without the short vowels that would disambiguate related word forms. In each case a keyword matcher is not slightly less effective — it is measuring something other than what you asked for.
Choosing a mode
| Mode | Best for | Trade-off |
|---|---|---|
keyword | Exact identifiers — tickers, case numbers, precise phrases you know appear verbatim. | Predictable and auditable, but only finds the phrasing you supplied. |
semantic | Concepts and themes, and anything a model wrote. Essential outside English. | Will surface related material you did not literally ask for. |
hybrid | Almost everything else. The sensible default. | Neither maximally precise nor maximally broad, which is usually right. |
The practical rule: if a person typed the query and knows the domain, keyword is defensible. If a model composed it, use hybrid — models paraphrase, and they paraphrase towards their own training distribution rather than towards newsroom style.
What semantic retrieval does not fix
It will not fix an unbounded query. Meaning-based matching always returns its best guesses, so a vague question over a deep archive produces a confident-looking result set with nothing anchoring it to the period you cared about. Bound every query with a time range; this matters more with semantic modes than with keyword, not less.
It also will not deduplicate. Forty outlets running the same wire story are forty relevant matches, and semantic retrieval will correctly return all of them. That is a different problem, solved by story grouping, and it is the single most common reason a well-built news pipeline still produces bad results.