Engineering Blog and Release Notes
Release announcements and engineering notes from the MDDB team: what shipped in each version, why we built it, and how the internals actually work.
-
The Duplicate Pages Your Crawler Cannot See
Three duplicate detectors — content hash, text overlap and embeddings — on one client site, with the thresholds measured rather than guessed.
-
What's New in MDDB 2.14
MDDB 2.14: the stateless MCP 2026-07-28 revision served beside 2025-11-25, asymmetric embeddings, retrievalMode honoured, and readable MCP resources.
-
Why Your RAG Returns the Wrong Document
A measured walk through RAG retrieval in MDDB: where semantic search fails, why no blend weight rescues it, and the one change that fixed it.
-
Is It Leaking? Read a Different Number
Why resident memory grows under sustained load in MDDB, which figure separates a cache from a leak, and a 45-minute measurement that settles it.
-
Flat Throughput Is a Diagnosis
How a constant 6.8 GB/s across every working-set size revealed that MDDB's vector comparison was instruction-bound, not memory-bound — and what fixed it.
-
What's New in MDDB 2.12
A practical overview of MDDB 2.12: search and RAG changes, measured performance gains, security fixes and the required upgrade steps.
-
What's New in MDDB 2.13
MDDB 2.13: native Windows testing, AVX2 vector maths, two data-loss fixes in restore, and one breaking change to user deletion.
-
The Collection Knows How to Be Searched
Per-collection retrieval profiles and response prompts in MDDB 2.12 — the collection carries how it should be searched and how its answers should be shaped.
-
What Breaks If I Change This?
Code symbol extraction and a connection graph in MDDB 2.12 — ask what depends on a CSS selector or a JS module, instead of grepping and hoping.
-
Choosing a Search Algorithm from the Collection
How the MDDB search advisor profiles a collection, recommends search settings and explains the limits of its vocabulary heuristic.
-
Diversifying Near-Duplicate Search Results
How MinHash-based diversity demotes near-duplicate search results in MDDB and changes a measured top-four result from one distinct answer to four.
-
Your WordPress just became AI-searchable: MDDB Sync
MDDB Sync mirrors your WordPress posts into MDDB for semantic search and RAG — and lets your AI agent publish drafts back. Setup takes ten minutes.
-
MDDB 2.11.4: tenants, better RAG, vectors that fit
Native multi-tenancy, parent/chunk/window retrieval, MMR diversification and disk-only quantized vectors - what shipped in MDDB 2.11.4, and why.