Skip to main content
Vector search finds the most similar vectors to your query using distance metrics instead of exact keyword matching. VectorAI DB compares vector embeddings to find content with similar meaning. When you submit a query, VectorAI DB compares it against stored vectors using a distance metric to measure similarity, ranks results by score, and returns the top matches. This approach works across diverse data types, including text, images, audio, and structured data. Common use cases include:
  • Semantic text search
  • Recommendation systems
  • Similar item discovery
  • Retrieval-Augmented Generation (RAG)
  • Image similarity
  • Audio matching

​
Search parameters

Configure search behavior with the following parameters: Adjust parameters based on your accuracy and speed requirements. Request only the data you need to minimize response size and network overhead.

​
Search results

Each search result contains the following components:

​
Always included

  • id: Vector identifier as an integer.
  • score: Similarity measure as a float.

​
Conditionally included

  • payload: Metadata dictionary when with_payload=True.
  • vector: Embedding array when with_vectors=True.

​
Understanding scores

Score interpretation varies by distance metric:
  • Cosine similarity uses higher scores to indicate greater similarity, with a range of negative one to one.
  • Euclidean distance uses lower scores to indicate greater similarity, with a range of zero to infinity.
  • Dot product uses higher scores to indicate greater similarity, with a range of negative infinity to positive infinity.
The distance metric setting at collection creation determines how VectorAI DB calculates scores. See distance metrics for detailed explanations.

​
Search performance optimization

Several strategies improve search performance and reduce latency.

​
Use filters effectively

Filters reduce the search space before similarity computation:
  • Apply restrictive filters to limit candidates.
  • Combine multiple conditions strategically.
  • Index frequently filtered fields for faster evaluation.
Pre-filtering is more efficient than post-filtering results because it reduces the number of similarity calculations.

​
Control result size

Request only the data you need:
  • Set limit to the minimum number of results required.
  • Set with_payload=False when metadata is unnecessary.
  • Set with_vectors=False unless raw embeddings are needed.
Smaller responses reduce network transfer time and improve perceived performance.

​
Tune HNSW parameters

Balance accuracy and speed with the ef parameter. Lower ef values provide faster search with reduced accuracy. Higher ef values provide better accuracy with increased latency. The default provides balanced performance for most use cases. Test different ef values with your data to find the optimal setting for your accuracy requirements.

​
Batch searches

For multiple queries, batch them together when possible. Batching reduces network round trips, improves overall throughput, and optimizes resource utilization. This is especially effective when processing many queries in sequence.

​
Task guides

The following guides cover search operations step by step.

​
Next steps

  • Filtering — Apply must, should, and must-not conditions to search.
  • Distance metrics — Understand how scores are calculated.
  • Indexing — Tune HNSW parameters for search performance.
  • Points — Manage the data being searched.