Skip to main content

​
ANN vs Exact Search

When searching for similar vectors, you have two fundamental choices: find the mathematically exact nearest neighbors, or find approximate nearest neighbors that are close enough. This page explains the tradeoff and why VecLabs - like most production vector databases - uses approximate search.
Exact search compares your query vector against every single stored vector and returns the mathematically precise top-K results. It is also called brute force search or flat search. Pros:
  • Perfect recall - you always get the true nearest neighbors
  • Simple to implement and reason about
  • No parameters to tune
Cons:
  • O(n) per query - every query requires examining every vector
  • At 1M vectors with 1536 dimensions: ~6 billion floating-point operations per query
  • Latency grows linearly with dataset size - 10x more vectors = 10x slower queries
  • Impractical above ~10K vectors for real-time applications
ANN search uses an index to skip most comparisons and find vectors that are very close to the true nearest neighbors, but not necessarily the exact ones. Pros:
  • Sublinear query time - grows very slowly with dataset size
  • Sub-millisecond to low-millisecond latency at millions of vectors
  • Tunable recall-speed tradeoff
Cons:
  • Not 100% accurate - might miss some true nearest neighbors
  • Requires building and maintaining an index
  • Index takes additional memory

​
The accuracy tradeoff in practice

ANN search is not as scary as “approximate” sounds. Consider: At 98% recall with top-10 results:
  • You get 9-10 of the true 10 nearest neighbors
  • Occasionally you might get result #11 instead of result #10
  • Your LLM or downstream system receives highly relevant context either way
For AI applications - RAG, agent memory, semantic search - this level of accuracy is indistinguishable from perfect. No user can tell the difference between the 10th and 11th most semantically relevant chunk. The only cases where exact search is truly necessary are:
  • Fraud detection where every false negative has a cost
  • Deduplication where you need 100% certainty
  • Legal or compliance applications requiring exact matching
For everything else in AI: use ANN.

​
VecLabs recall numbers

VecLabs HNSW with default parameters (M=16, ef_construction=200):
Dataset sizeRecall @ top-10
10K vectors~99.5%
100K vectors~98.2%
1M vectors~97.1%
These are recall numbers at the default ef_search setting. Recall can be increased by raising ef_search at the cost of slightly higher query latency.

​
Exact search in VecLabs

VecLabs does not currently expose an exact search mode. The HNSW index is always used. If you have a dataset under 1000 vectors and need exact results, the performance difference between ANN and exact search is negligible - HNSW at that size is effectively exact anyway due to the small world property of the graph.