Skip to main content
Kora provides approximate nearest neighbor search using HNSW (Hierarchical Navigable Small World) indexes. Vectors support cosine similarity, L2 distance, and inner product metrics.

​
VECSET

Store a vector in the index. Syntax 
VECSET key dim v1 v2 v3 ...
​
key
string
required
The key to store the vector under
​
dim
integer
required
The dimensionality of the vector
​
v1 v2 ...
float[]
required
The vector components (must match dim count)
Return value
​
result
string
OK if the vector was stored successfully
Examples 
redis-cli> VECSET vec:1 3 1.0 0.5 0.8
OK
redis-cli> VECSET vec:2 3 0.9 0.6 0.7
OK
redis-cli> VECSET embedding:product:123 128 0.1 0.2 0.3 ...
OK
Time complexity: O(M log N) where N is the number of vectors and M is the connectivity parameter Notes:
  • Vectors are automatically normalized for cosine similarity
  • All vectors in an index must have the same dimensionality
  • The first VECSET determines the dimension for subsequent inserts

​
VECQUERY

Find the K nearest neighbors to a query vector. Syntax 
VECQUERY key k v1 v2 v3 ...
​
key
string
required
The index key pattern or base key
​
k
integer
required
Number of nearest neighbors to return
​
v1 v2 ...
float[]
required
The query vector components
Return value
​
results
array
Array of (key, distance) pairs, sorted by distance
Examples 
redis-cli> VECQUERY index 5 1.0 0.5 0.8
1) "vec:2"
2) "0.0523"
3) "vec:1"
4) "0.1247"
5) "vec:5"
6) "0.2891"
7) "vec:3"
8) "0.3456"
9) "vec:7"
10) "0.4123"
Time complexity: O(ef * log N) where ef is the search width parameter and N is index size Notes:
  • Returns up to k results (may be fewer if index has fewer vectors)
  • Distances are cosine distances (lower is more similar)
  • Query vector is automatically normalized

​
VECDEL

Delete a vector from the index. Syntax 
VECDEL key
​
key
string
required
The key of the vector to delete
Return value
​
result
integer
1 if the vector was deleted, 0 if it did not exist
Examples 
redis-cli> VECDEL vec:1
(integer) 1
redis-cli> VECDEL vec:99
(integer) 0
Time complexity: O(M log N) where N is the number of vectors and M is the connectivity parameter

​
Configuration

​
HNSW Parameters

The HNSW index is configured with the following default parameters:
  • M (connectivity): 16 - Number of bi-directional links per node
  • ef_construction: 200 - Search width during index construction
  • ef_search: 100 - Search width during query (configurable per query)

​
Distance Metrics

Kora supports three distance metrics:
  1. Cosine Similarity (default)
    • Measures angular similarity between vectors
    • Range: [0, 2] where 0 is identical
    • Automatically normalizes vectors
  2. L2 Distance (Euclidean)
    • Measures straight-line distance
    • Range: [0, ∞)
    • No normalization applied
  3. Inner Product
    • Dot product similarity
    • Range: (-∞, ∞)
    • Higher values indicate greater similarity

​
Performance Characteristics

​
Index Build

  • Memory usage: ~(4 * dim + 8 * M) bytes per vector
  • Build time: O(N * M * log N) for N vectors
  • Suitable for millions of vectors with hundreds of dimensions

​
Query Performance

  • Typical recall@10: Greater than 95% with default parameters
  • Query latency: Less than 1ms for 100K vectors, less than 10ms for 1M vectors
  • Scalable to 10M+ vectors per shard

​
Tuning Guidelines

For higher recall:
  • Increase M (e.g., 32 or 64)
  • Increase ef_construction (e.g., 400)
  • Trade-off: larger memory footprint, slower indexing
For faster queries:
  • Decrease ef_search
  • Trade-off: lower recall
For faster indexing:
  • Decrease M and ef_construction
  • Trade-off: lower recall

​
Use Cases

# Store document embeddings
VECSET doc:1 768 0.123 -0.456 ...
VECSET doc:2 768 0.234 -0.567 ...

# Search for similar documents
VECQUERY index 10 0.111 -0.444 ...

​
Image Similarity

# Store image feature vectors
VECSET img:cat.jpg 512 0.91 0.23 ...
VECSET img:dog.jpg 512 0.87 0.31 ...

# Find similar images
VECQUERY index 5 0.90 0.25 ...

​
Recommendation Systems

# Store user/item embeddings
VECSET user:alice 64 0.5 0.3 ...
VECSET item:book:123 64 0.4 0.6 ...

# Find similar users or items
VECQUERY index 20 0.45 0.35 ...

​
Implementation Notes

  • HNSW index is implemented per-shard with no cross-shard coordination
  • Vectors are stored in-memory for fast access
  • Index structure is optimized for read-heavy workloads
  • Supports concurrent queries without locks
  • Index is not persisted by default (rebuild on restart)

​
Comparison with Redis

Kora’s vector search is similar to RediSearch’s vector capabilities but with key differences:
  • Kora uses pure HNSW (no flat index fallback)
  • Lower memory overhead per vector
  • Better multi-threaded query performance due to shard-affinity
  • Simpler command interface (no FT. prefix)
  • No persistence of vector index (in-memory only)

Build docs developers (and LLMs) love

Get started for freeTalk to us