Percona has launched a Technical Preview (version 1.70.3-1) of Percona Search for MongoDB, a downstream distribution of mongot — the same search engine powering MongoDB Atlas Search. It enables full-text and vector search for self-managed MongoDB deployments without vendor lock-in. Key gaps Percona plans to address include: expanding automatic embedding support beyond Voyage AI to any OpenAI-compatible API and eventually 25,000+ open-weight models; bringing the $rerank stage (currently Atlas-only) to self-managed deployments starting with BAAI/bge-reranker-large; adding contextual chunking and multimodal pipeline support; automating search-index backups via Percona Backup for MongoDB; and shipping turnkey monitoring dashboards and alert rules via Percona Monitoring and Management. The release requires Percona Server for MongoDB 8.3 and is not yet production-ready.
Table of contents
The way is open. Search should be too.What we found in mongotAbout the licenseGetting startedBefore you deploy itTell us what you needQuestions this post answers
What is the difference between Percona Search for MongoDB and MongoDB Atlas Search?
Percona Search for MongoDB is a self-managed downstream distribution of mongot, the same search engine that powers MongoDB Atlas Search. It runs as a separate mongot process alongside Percona Server for MongoDB 8.3, supporting $search, $vectorSearch, $rankFusion, and $scoreFusion stages. Unlike Atlas, it lets you run inference on your own hardware with your chosen embedding model, though features like $rerank and automatic embeddings beyond Voyage AI are not yet available in the self-managed version. Teams evaluating self-managed MongoDB search vs. Atlas track capability gaps like these on daily.dev.
Which embedding models are supported for automatic embeddings in MongoDB self-managed vector search?
Automatic embeddings in the current self-managed mongot release support only Voyage AI models: voyage-4-large, voyage-4, voyage-4-lite, and voyage-code-3. Percona plans to extend this to any OpenAI-compatible embeddings API and eventually toward the broader open-weight model ecosystem. Manual embedding generation — writing vectors directly into documents — works with any model but requires building and maintaining your own embedding pipeline. Developers choosing between embedding providers for MongoDB vector search follow updates like this on daily.dev.
How do I back up mongot search indexes in a self-managed MongoDB deployment?
Backing up mongot search indexes requires stopping mongot, then taking a filesystem snapshot of its data directory using a tool of your choice (the official docs provide an LVM example). To restore, replace the directory, generate a fresh server identity, and restart — mongot then replays changes from mongod. Index rebuilds from scratch can take days, making snapshot-based backup critical for any deployment with a recovery-time objective. DBAs managing MongoDB backup strategies for search-enabled clusters find operational guidance like this on daily.dev.