Concept Contrast Fable

Hybrid search and dense retrieval

Conclusion first:Dense retrieval uses vectors to find semantically similar content, and hybrid search is a combined strategy that pairs dense retrieval with keyword matching. Treating the combination as a component will miss the other half that actually makes the whole process work.

Dense retrieval uses vectors to find semantically similar content, and hybrid search is a combined strategy that pairs dense retrieval with keyword matching. Treating the combination as a component will miss the other half that actually makes the whole process work.

Fable

Two combs, one only smooths the fur

A new retrieval clerk named Gu arrived at the North City Archive. The archive held tens of thousands of documents, and when readers came to look up materials, Sister Gu searched with a keyword index book. The index book was arranged by entry, so when a reader said "water conservancy," she turned to the water section and found all documents containing "water conservancy." It was fast, but when a reader asked about "the historical evolution of canal irrigation," there was no "canal" entry in the index book, so she could only guess a few synonyms from experience and search.

The curator brought back a "mirror of meaning" from elsewhere. This mirror could turn a passage of text into a string of numbers, and documents whose numbers were close in the mirror had similar meanings. Sister Gu tried using it to search for "canal irrigation," and the mirror directly returned a batch of documents about irrigation engineering and canal construction, even though the words "canal irrigation" did not appear in them. Sister Gu thought this thing was amazing, so she put the index book into a drawer and used only the mirror.

A month later, the problem surfaced. A reader wanted to look up "the Yellow River breach in the third year of Hongwu," but the mirror returned a pile of modern papers on Yellow River hydrological management, missing the original memorial from the third year of Hongwu entirely—because the memorial said "the river overflowed," which was not a near neighbor of "breach" in the mirror. Sister Gu searched for a long time and came back empty-handed, and the reader was very dissatisfied.

Sister Gu pulled the index book out of the drawer and looked up the three words "Hongwu," "Yellow River" and "water" separately. Cross-comparing the three indexes, the memorial from the third year of Hongwu immediately surfaced — keyword matching is naturally sensitive to proper nouns, precise years, personal names and place names, which is something a mirror cannot do.

Sister Gu changed her workflow from then on: every time she looked something up, she first used the index book to pull a batch of exact-match candidate documents, then used the meaning mirror to add a batch from semantically similar directions. After deduplicating and merging the two batches, she sent them together into the reading room. She called this combined workflow "hand-in-hand search."

A new intern saw Sister Gu's process and asked her: isn't this just the meaning mirror? Sister Gu shook her head: the mirror only handles the semantic path; the joint check also has the index book path. What you see is the same final result, but behind it two pipelines converge.

Sister Gu later summed it up: the meaning mirror is good at finding "similar in meaning," while the index book is good at finding "literal and exact." In everyday retrieval, the literal-and-exact path is often ignored, until you run into proper nouns, code, numbers, or dates, and discover that the index book is the only reliable entry point.

Later the archive room locked the two tools into the same cabinet: the meaning mirror handled broad reading, and the index book handled precise hits. Only together could it be called archive search; neither alone was enough.

Dense Retrieval

Vector embeddings map queries and documents into the same semantic space, and relevance is measured by vector distance. They are good at capturing synonymous expressions and semantic associations, but not sensitive enough to proper nouns and exact matches.

Hybrid Search

A combined strategy that fuses the results of dense retrieval and sparse retrieval (keyword matching). It leverages both semantic understanding and exact matching signals, usually merging the final results through weighting or reranking.

Story Mapping

Keyword index book: sparse retrieval (BM25, etc.), exact term matching.
Mirror of Meaning: dense retrieval, vector semantic matching.
Hybrid search: hybrid search, fusing results from two paths.
Missed detection in a Hongwu third-year memorial: the recall blind spot of dense retrieval for proper nouns and precise expressions.
Intern confusion: hybrid search ≠ dense retrieval; the former includes the latter but is not limited to it.

In the project

When you hear someone say "our search uses Dense Retrieval," ask one follow-up question: is keyword matching also fused in? If the answer is "only vector retrieval," then precise queries such as proper nouns, dates and code numbers will most likely run into problems. Hybrid search cannot be treated as an upgraded version of dense retrieval; it is a combination of the dense and sparse paths. The weight of each path and the fusion method (linear weighting or reranking with a Cross-Encoder) are the real decision points that affect performance.

Hybrid search Dense retrieval