Keyword vs semantic search

How keyword search works

Keyword search, also called lexical or full-text search, builds an index of which words appear in which documents. A query is matched against that index, and results are ranked by how often and how prominently the words appear, commonly with a scoring method called BM25. Good implementations handle plurals and word endings (“run,” “running”), but they still need the actual words to be present.

Keyword search is also predictable: you can see exactly why a result matched, because the word is right there.

How semantic search works

Semantic search converts both the query and the content into embeddings and returns the content whose meaning is closest. It doesn't need shared words at all. “Laptop battery drains fast” can match “my notebook won't hold a charge.”

Side by side

Imagine searching recorded customer calls for complaints about delivery delays:

• Query “late delivery”, keyword: finds calls where someone said “late delivery.” Misses “the package took three weeks.”
• Same query, semantic: finds both, and may also surface a call about a late refund, which is related but not what you wanted.
• Query “order 48213”, keyword: finds the one call where that number was read out.
• Same query, semantic: may return any call that mentions an order number, because the digits carry little meaning.

Why hybrid search is common

Because the two approaches fail differently, many search systems run both and merge the rankings, a setup called hybrid search. Keyword matching guarantees that exact names, codes and quotes are found; semantic matching catches paraphrases and vague memories. Some systems add a reranker that reads the query and each candidate together to decide the final order. As a rule of thumb: if you remember the exact words, keyword search is enough. If you remember the idea, you need semantic search.

For spoken media there is one more wrinkle. Transcripts contain recognition errors, so a keyword search for a name can fail simply because the transcript spelled it differently. Semantic search is somewhat more forgiving of this, and phonetic matching, which compares how words sound rather than how they are spelled, closes more of the gap.

In MediaFind

MediaFind searches your transcripts by meaning, so a loose description finds paraphrases. For the cases where meaning alone falls short, it adds a phonetic sounds-like search for words the transcript may have misheard and an entity search that finds every time a person, place or organization came up. It also covers what neither approach can, such as on-screen text, visuals and speakers. Everything runs on your computer. Read how MediaFind's search channels fit together for more detail.

Frequently asked questions

Is semantic search always better than keyword search?

No. For exact names, IDs, codes and quotes, keyword search is usually more reliable. Semantic search is better when you remember the idea but not the wording.

What is hybrid search?

Hybrid search runs keyword and semantic search together and merges their results, so you get exact matches and paraphrases in one list.

Does Ctrl+F count as keyword search?

Yes, in its simplest form. It finds exact character matches in one document, with no ranking and no handling of word variants.