Core Technology

A conventional vector database loads the entire HNSW graph into RAM, which works at small scale but becomes prohibitively expensive at billions of vectors. Endee's Layered Memory Architecture separates the vector graph into two tiers.

Base-layer graph nodes are persisted on disk using memory-mapped files. The OS page cache acts as a warm tier, frequently accessed nodes stay resident in memory naturally, without explicit management. The upper layers of the HNSW graph, the small-world navigational links that determine routing quality, are held in a configurable in-memory cache.

The result: RAM consumption drops by orders of magnitude compared to a fully in-memory index, while recall remains stable because the high-connectivity upper layers stay hot. Latency stays low because base-layer access via mmap is faster than a network round-trip to a distributed cache.

Float32 vectors are 4 bytes per dimension. At 768 dimensions and 100M vectors, that's 300GB of raw vector storage, before indexing overhead. Standard INT8 quantization halves or quarters this, but introduces precision loss that degrades recall.

Endee's Int8e (Int8 enhanced) format is a proprietary encoding that reduces vector footprint by 4x compared to float32 while preserving retrieval accuracy through three mechanisms: per-vector dynamic scale calibration, selective precision retention for dimensions with high variance, and a correction step applied during distance computation.

Int8e is an enterprise-only format. Open-source builds support standard INT8 and INT16 with dynamic per-vector scaling, which already deliver significant compression at high accuracy.

Applying a metadata filter to an HNSW search creates a fundamental tension: the graph was built on the full dataset, but the query must return only the filtered subset. Naive approaches prune edges aggressively to enforce filters early, which disconnects the graph and degrades recall dramatically, particularly for selective filters.

Endee's Progressive Filtering applies constraints incrementally across the graph traversal. Rather than blocking graph edges at the outset, the algorithm progressively tightens the candidate set while preserving enough connectivity to navigate toward globally optimal results. A final check enforces hard filter constraints on the output set.

The result is recall that degrades gracefully with filter selectivity, rather than collapsing. Highly selective filters (0.1% matching) still return accurate top-k results that single-pass approaches miss.

Standard vector databases require decrypting your data to run search. The server receives plaintext vectors, builds the graph on plaintext, and executes queries against plaintext. For regulated industries, finance, healthcare, defense, this is architecturally incompatible with data sovereignty requirements.

Endee's Queryable Encryption enables ANN similarity search directly on encrypted vector representations. Encryption and decryption happen exclusively on the client side. The Endee server stores and processes only ciphertext throughout the entire data lifecycle: ingest, indexing, and query execution.

The encryption scheme is designed to preserve approximate distance ordering in the encrypted space, enabling graph traversal without decryption. The server learns nothing about the content of the vectors it processes.