What is hybrid search?

How hybrid search works

A hybrid system sends the query down two paths at once. The keyword path looks for the actual words, usually with a ranking formula such as BM25 that rewards rare terms appearing often in a passage. The semantic path turns the query into an embedding and finds passages with nearby meaning. Each path returns its own ranked list, and a fusion step combines them.

The two lists cannot simply be added, because their scores are on different scales. A common fix is reciprocal rank fusion (RRF), which ignores raw scores and uses positions: each result earns 1/(k + rank) from every list it appears in, with k often set to 60, and the totals are sorted. A passage ranked well by both paths rises to the top. The alternative is to normalize both scores and blend them with a weight, which gives more control but needs tuning.

An example

Suppose your meeting recordings include one where a colleague says “the Henderson invoice, INV-20931, is still disputed” and another where someone says “that client still hasn't paid the bill we argued about.” Search for INV-20931 and keyword matching finds the first instantly, while semantic search may drift to any talk about invoices. Search for the unpaid bill we argued over and semantic search finds the second, while keyword matching finds nothing useful. Hybrid search returns the right result for both queries without you deciding which mode to use.

Pitfalls

• Weights that favor one side. Blend too much toward semantic and exact codes sink; too much toward keywords and paraphrases vanish.
• Duplicate hits. The same moment found by both paths should merge into one result, not appear twice.
• Transcription errors hurt the keyword side. If a name was transcribed wrong, keyword search cannot find it, though semantic or phonetic matching might.
• Fusion does not fix a bad first stage. If neither path retrieves the right passage, merging will not surface it. A reranking pass reorders candidates but cannot add missing ones.

In MediaFind

MediaFind runs both exact and semantic search on your library and fuses the results, so a part number and a half-remembered description both work from the same search box. Beyond text, the same fused results can draw on other ways of searching, such as on-screen text, speakers, visual scenes and a phonetic sounds-like mode for words that were transcribed a little wrong. Every result links to an exact timestamp. Read more in keyword vs semantic search and the engineering note on result fusion.

Frequently asked questions

Is hybrid search better than semantic search alone?

For most real collections it is more reliable, because it covers exact names, numbers and codes that semantic search tends to blur, while keeping paraphrase matching.

What is reciprocal rank fusion?

A way of merging ranked lists using only each item's position. Every list contributes 1 divided by (a constant plus the rank), and items are sorted by the total.

Does hybrid search need two separate indexes?

Usually yes: a keyword index over the text and a vector index over the embeddings, queried in parallel.