Use Case

Recommendations — nearest neighbours at production scale.

Embed users, items, and interactions into the same vector space. Use HNSW nearest-neighbour retrieval to power product recommendations, content feeds, and user matching — with the storage economics of object storage and the latency of a purpose-built index.

The challenge

Recommendation systems need fast, scalable ANN.

Two-stage recommenders retrieve candidates via approximate nearest-neighbour lookup, then re-rank. The ANN retrieval step needs to be fast, scalable, and cheap — properties in-memory vector databases struggle to deliver at catalog scale.

The solution

ManyVector as the candidate retrieval layer.

Store user and item embeddings in ManyVector. Retrieve the top-k nearest items for a given user vector in under 10ms. Namespaces per user segment, per A/B variant, or per content category — all from your own bucket.

Patterns

Common recommendation architectures.

Item-to-item

Similar items

Embed item A and retrieve its nearest neighbours from the item namespace. Powers "customers also bought," "similar products," and "you may also like" patterns.

User-to-item

Personalized feed

Aggregate user interaction embeddings into a user vector. Retrieve nearest items from the catalog namespace. Personalization at scale without co-matrix factorization overhead.

Session-based

In-session context

Embed the current session sequence and retrieve items relevant to what the user is doing right now. Works for anonymous users with no interaction history.

By the numbers

Real-time personalization at scale.

<10ms

recommendation serving

Any

item type

Real-time

personalization

Filters

for business rules

FAQ

Common questions

How do I compute user embeddings for personalization?

Average the embeddings of items a user has interacted with, or train a user tower on your interaction data. ManyVector stores and queries these like any other vector.

Can I filter recommendations by category or availability?

Yes. Add category, price, availability, or any attribute as metadata. Pass filter conditions alongside your vector query to enforce business rules.

How do I handle cold-start for new items with no history?

Use content-based embeddings — image features, text descriptions — for new items. These work immediately without any interaction data.

Can ManyVector serve real-time recommendation APIs?

Yes. The query API is designed for sub-10ms p50 serving. Many production recommendation systems serve directly from ManyVector behind an API gateway.

Recommendations at scale on your own bucket.