105 lines
2.3 KiB
Markdown
105 lines
2.3 KiB
Markdown
|
# Search and Vector Query Engines
|
||
|
|
|
||
|
|
A reference for how search-oriented engines differ from classical relational ones.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Short answer
|
||
|
|
|
||
|
|
Search and vector engines are still query engines, but their core operators, indexes, and ranking models differ from standard relational SQL systems.
|
||
|
|
|
||
|
|
They are often built around retrieval rather than exact relational transformation.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Search engines
|
||
|
|
|
||
|
|
Traditional search engines center on:
|
||
|
|
|
||
|
|
- inverted indexes
|
||
|
|
- document retrieval
|
||
|
|
- ranking and scoring
|
||
|
|
- boolean and text queries
|
||
|
|
|
||
|
|
The core question is often not just "which rows satisfy a predicate?" but "which documents are most relevant?"
|
||
|
|
|
||
|
|
That makes ranking a first-class part of execution.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Vector engines
|
||
|
|
|
||
|
|
Vector engines center on:
|
||
|
|
|
||
|
|
- embeddings
|
||
|
|
- nearest-neighbor search
|
||
|
|
- approximate similarity retrieval
|
||
|
|
- metadata filtering
|
||
|
|
|
||
|
|
The core operation is often "find the closest vectors to this query vector," which is different from equality joins or exact predicate evaluation.
|
||
|
|
|
||
|
|
Approximation is often acceptable because speed matters and exact nearest-neighbor search can be too expensive at scale.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Hybrid systems
|
||
|
|
|
||
|
|
Many modern systems combine:
|
||
|
|
|
||
|
|
- lexical search
|
||
|
|
- vector search
|
||
|
|
- metadata filters
|
||
|
|
- reranking
|
||
|
|
|
||
|
|
This means the engine may need to merge several candidate-generation and scoring paths into one final result.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## How these engines differ from relational ones
|
||
|
|
|
||
|
|
Search and vector engines often emphasize:
|
||
|
|
|
||
|
|
- ranking
|
||
|
|
- retrieval quality
|
||
|
|
- approximate indexing
|
||
|
|
- candidate generation
|
||
|
|
- top-k execution
|
||
|
|
|
||
|
|
Relational engines more often emphasize:
|
||
|
|
|
||
|
|
- exact semantics
|
||
|
|
- joins
|
||
|
|
- aggregations
|
||
|
|
- transactional or analytical correctness
|
||
|
|
|
||
|
|
The difference is not that one has a query engine and the other does not. The difference is what the engine is optimizing for.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Example systems
|
||
|
|
|
||
|
|
Examples of search and vector query engines include:
|
||
|
|
|
||
|
|
- Lucene
|
||
|
|
- Vespa
|
||
|
|
- Weaviate
|
||
|
|
- Qdrant
|
||
|
|
|
||
|
|
They differ in packaging and scope, but all have real planning and execution concerns around retrieval.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Practical mental model
|
||
|
|
|
||
|
|
Relational engines are often about exact transformation of structured data.
|
||
|
|
|
||
|
|
Search and vector engines are often about efficient retrieval and ranking over high-dimensional or text-heavy data.
|
||
|
|
|
||
|
|
That is the cleanest conceptual distinction.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## Changelog
|
||
|
|
|
||
|
|
* **April 1, 2026** -- First version created.
|