Embedding search gets its index

OrionAI 2.44.2Released 9 November 2026

The dataset-embedding table now carries an HNSW vector index. Similarity search (related datasets, answer-engine lookups) no longer scans every embedding — milliseconds instead of tens of seconds, and the 2026-11-09 database saturation that took the platform down cannot recur on any database created from this schema.

Changes since 2.44.1.

#Fixed

  • Similarity search now uses an HNSW vector index. index_embeddings shipped without any vector index, so every similarity query brute-forced all ~64,000 embeddings; a burst of concurrent searches saturated the vector database and the platform database (2026-11-09 outage). The index is now created with the schema (CREATE INDEX IF NOT EXISTS), so fresh databases get it from day one and existing ones keep the index already created in production.

#Notes

  • Index builds allocate their graph in shared memory — the pgvector service in the production compose file now carries shm_size: 2g (the Docker default of 64 MB made the first build crawl).