E-commerce search
Shoppers type what they want in their own words. The catalog is searched by meaning and filtered to the category and the shop they are in.
AI and LLMRetrieval
Vector Similarity Search
Find the nearest items by meaning, filtered by tag, inside the same engine that holds the items themselves.
- indexed
- HNSW
- average search
- O(log N)
- distance metrics
- 3
Step 1 of 6
Create the index
Dimensions and metric are fixed per index. HNSW parameters can be tuned when needed.
Example sessionplaying
# Create the index
hotkv> VSIM.CREATEINDEX products 3 METRIC COSINE
OK
An animated walkthrough with real HotKV commands and example replies. Click any step to jump to it.
The problem
Semantic search, "more like this" and deduplication all come down to one question. Which stored items are closest to this one? Answering it usually means a dedicated vector database next to your main store, and every product update has to be written to both.
How HotKV handles it
HotKV keeps HNSW vector indexes in the engine. Product data lives in ordinary keys, the vectors are indexed under the same key names, and tags on those keys let a search stay inside a category or a tenant. One write path, one system to run.
Step by step
What happens, command by command
- 01
Create the index
Dimensions and metric are fixed per index. HNSW parameters can be tuned when needed.
VSIM.CREATEINDEX products 3 METRIC COSINE - 02
Store the product
The product's data is a normal hash, under the key name the index will use.
HSET sku:1 name "Trail runner" price 89 - 03
Tag it
Tags on the key let a vector search stay inside a category.
TAG.SET sku:1 shoes running - 04
Index its vector
The vector goes into the index under the same key name.
VSIM.SET products sku:1 0.12 0.80 0.31 - 05
Search by meaning, within a tag
The query text is embedded, and the nearest products that carry the tag come back with their distances.
VSIM.SEARCH products 3 0.10 0.82 0.29 FILTER shoes - 06
Fetch what to show
The results are key names, so the product data is one read away in the same server.
HGETALL sku:1
Where it fits
Real-world scenarios
Duplicate detection
New support tickets or listings are compared with existing ones, and near duplicates are merged before a human sees them.
Image and media similarity
Embeddings from an image model find visually similar photos or clips across a library.
Built in
What you get
HNSW indexes
Approximate nearest-neighbour search that stays fast on large indexes, O(log N) on average.
VSIM.CREATEINDEXVSIM.SEARCHThree metrics
Cosine, L2 and inner product, chosen per index.
METRICTag filters
Restrict a search to keys that carry every given tag.
TAG.SETFILTERPersistent
Indexes are kept in your snapshots and AOF rewrites like the rest of your data.
VSIM.INFOCommon questions
Does HotKV generate the embeddings?
No. Create vectors with the embedding model you prefer and store them with VSIM.SET.
How is this different from RAG pipelines?
VSIM is the general vector index. RAG pipelines build on the same idea and add chunking, metadata and hybrid keyword search for documents.
Which plans include it?
Vector search and tags are core commands and work on every plan, with no license needed.
Related use cases
Try it on your own workload
HotKV speaks the Redis protocol, so you can point an existing client at it and run these same commands.