Use Case

    AI Recommendations with Endee

    Serve real-time product, content, or user recommendations at scale. One API call returns personalised results in milliseconds.

    Item EmbeddingsUser EmbeddingsReal-timeBusiness Rule FiltersBillion-scaleCollaborative Filtering

    Capabilities

    Built for personalization at scale

    Item and User Embeddings

    Store product, content, or user vectors from any embedding model. Retrieve the k-nearest items to a user vector for instant personalised recommendations. Supports collaborative filtering, content-based filtering, and hybrid approaches in the same index.

    Real-time Under 10ms

    Serve recommendations in a single API call at sub-10ms latency. No batch jobs, no precomputed tables that go stale overnight. Recommendations reflect the latest item catalog and user signal because the vector index is always live and current.

    Scales to One Billion Items

    Endee scales to one billion vectors on a single node using adaptive quantization. INT8 reduces memory by 75% with minimal recall impact, so a full billion-product catalog fits in 128 GB RAM without sharding or cluster management.

    Business Rule Filters

    Combine vector similarity with hard constraints such as in-stock status, geo-region, age rating, and price band. Filters are applied inside the ANN search graph for zero overhead. Results are both semantically relevant and compliant with business rules.

    Multi-strategy Personalization

    Mix collaborative, content-based, and session-based signals in one pipeline. Store item vectors, user preference vectors, and session embedding vectors in a single Endee index. Query against any of them independently or fuse the results at application level.

    Incremental Updates

    Add or update item vectors without full re-indexing. New products appear in recommendation results immediately after insert. Update user preference vectors in real time as behavior signals arrive, keeping personalization fresh without pipeline restarts.

    Process

    How it works

    1

    Embed items and users

    Generate item embeddings from product descriptions, images, or interaction data. Generate user embeddings from purchase history, clicks, or explicit ratings. Any model that produces fixed-dimension dense vectors works with Endee.

    2

    Index with metadata

    Insert all item vectors into Endee alongside structured metadata: category, price, stock status, region, and any custom attributes. Update vectors incrementally as catalog data changes without full re-indexing.

    3

    Retrieve at request time

    At request time, fetch the user embedding and call Endee with business-rule filters. Endee returns the top-k recommended items in milliseconds. No cache warming, no precomputed tables. Personalization is always fresh.

    In Practice

    What teams build with recommendations

    E-commerce "You May Also Like"

    Retrieve items nearest to what the user has browsed or purchased, filtered by stock and region.

    Streaming Content Recommendations

    Surface the next show or track using embeddings from watch history and audio/text content vectors.

    News Feed Personalization

    Rank articles by semantic similarity to a user's reading history, updated in real time as they read.

    B2B Product Matching

    Match supplier catalogs to buyer needs by comparing product specification embeddings.