Apache OpenNLP vs Weaviate

Weaviate 1.39 is the current stable release, having GA'd the Boost API and MMR diversity selection while introducing an experimental Search REST API. Two major storage advances landed in rapid succession: HFresh, a disk-based vector index that keeps vectors off the heap entirely, and 4-bit Rotational Quantization, which compresses stored vectors with minimal accuracy loss. Above the storage layer, Engram — a named agent memory product — is the most significant architectural addition: it lets builders configure extraction topics, scopes, and retrieval modes as first-class settings rather than building memory pipelines by hand.

Weaviate is executing a two-layer expansion: at the bottom, making the vector store cheaper and more flexible (quantization, disk-based indexing, query profiling); at the top, building agent-native abstractions that make Weaviate more than a search backend (Engram memory, effort tiers, Search REST API). The direction has shifted from 'fast vector database' toward 'infrastructure for AI agent memory and retrieval systems.' The consistent release of deep technical content alongside product updates suggests the team is deliberately targeting developers building production agent systems, not just evaluating vector databases.

Engram moving from guide to GA release is the most predictable next step. The experimental Search REST API, introduced in 1.39, is also positioned to stabilize — and the growing late-interaction retrieval work (multi-vector for PDFs and charts) looks like the foundation of a more formal multi-modal retrieval product rather than staying at the technique level.