MDDB bloghttps://mddb.tradik.com/2026-08-24T19:46:12ZIs It Leaking? Read a Different Numberhttps://mddb.tradik.com/blog/rss-is-the-wrong-number/2026-08-24T00:00:00Z2026-08-24T19:46:12ZA report said MDDB grew from 42 MB to 153 MB under load. The observation was probably right and the conclusion wrong — and the difference is one field you already have.Flat Throughput Is a Diagnosishttps://mddb.tradik.com/blog/the-loop-that-was-not-waiting-for-memory/2026-08-24T00:00:00Z2026-08-24T19:46:12ZTwo optimisations to MDDB's vector loop predicted 33% and delivered 1.5%. The benchmark had been saying why for months, in a column nobody read as a diagnosis.What's New in MDDB 2.12https://mddb.tradik.com/blog/whats-new-in-mddb-2-12/2026-08-24T00:00:00Z2026-08-24T19:46:12ZMDDB 2.12 adds code-aware retrieval, collection search profiles, faster bulk indexing, safer defaults and clearer production diagnostics.What's New in MDDB 2.13https://mddb.tradik.com/blog/whats-new-in-mddb-2-13/2026-08-24T00:00:00Z2026-08-24T19:46:12ZMDDB 2.13 runs its test suite on Windows, makes vector comparison 3.7–6.3× faster on amd64, and fixes two ways a restore could destroy the database it was restoring.The Collection Knows How to Be Searchedhttps://mddb.tradik.com/blog/the-collection-knows-how-to-be-searched/2026-08-23T00:00:00Z2026-08-24T19:46:12ZRetrieval settings and answer formatting used to live in every client. In MDDB 2.12 they live with the data, so a new consumer gets them right without being told.What Breaks If I Change This?https://mddb.tradik.com/blog/what-breaks-if-i-change-this/2026-08-23T00:00:00Z2026-08-24T19:46:12ZMDDB 2.12 indexes what a code file declares, uses and imports, then answers the relational questions grep cannot: what depends on this selector, which pages load this script, what is nobody using any more.Choosing a Search Algorithm from the Collectionhttps://mddb.tradik.com/blog/which-algorithm-should-i-use/2026-08-23T00:00:00Z2026-08-24T19:46:12ZMDDB now profiles a collection and recommends search settings from its document length, vocabulary, vector coverage and extracted symbols.Diversifying Near-Duplicate Search Resultshttps://mddb.tradik.com/blog/your-vector-search-rewards-duplicates/2026-08-23T00:00:00Z2026-08-24T19:46:12ZA runbook copied for four environments can occupy four result slots. MDDB uses MinHash overlap to reduce that repetition after ranking.Your WordPress just became AI-searchable: MDDB Synchttps://mddb.tradik.com/blog/wordpress-plugin/2026-08-02T00:00:00Z2026-08-24T19:46:12ZThere's a decade of knowledge sitting in your WordPress. Product announcements,
how-to guides, changelogs, that one post from 2019 that still answers half of
your support tickets. And when you plug an AI agent into your company, the
first thing it asks is, effectively: where's the knowledge?
The MDDMDDB 2.11.4: tenants, better RAG, vectors that fithttps://mddb.tradik.com/blog/mddb-2-11-4/2026-07-31T00:00:00Z2026-08-24T19:46:12ZWe spent the last cycle reading a lot of feedback — not just about MDDB, but
about what people using vector databases in general keep running into. A few
themes came up again and again: "I want to put multiple customers on one
instance without building isolation myself", "my RAG answe