From Qdrant to MariaDB Vector: Atarim's RAG in a single database
Atarim replaced Qdrant with MariaDB Vector for its RAG system: filtered queries in production in 10.3 ms, one fewer vendor to manage.
UptimeMag editorial team · 16 November 2026 · 3 min read

Atarim, an agency that uses AI agents for client website maintenance, has removed Qdrant from its RAG (retrieval-augmented generation) system and moved vectors and metadata into MariaDB. The team describes the switch as an almost direct replacement, made to integrate data more easily, keep it under a single access control, and manage one fewer vendor. For anyone administering databases on VPS or in the cloud, it's a concrete case study of what changes when switching from a dedicated vector database to a VECTOR column inside an RDBMS already in use.
Why MariaDB 11.8
Atarim runs its RAG on AWS, with MariaDB 11.8 on Amazon RDS. The version choice is deliberate: it's the first LTS release of MariaDB with MariaDB Vector, the similarity search function built into the server without extensions, and the first available with this capability on Amazon RDS for MariaDB.
The production schema stores embeddings (1,536 dimensions, OpenAI text-embedding-3-small model) and JSON metadata in the same table, with a VECTOR INDEX based on HNSW (M=16, cosine distance) and virtual columns indexed for fields used in filters. The application runs Laravel 12 with the native MariaDB driver.
Numbers measured by Atarim
In production, on a db.t4g.micro instance with 9,708 vectors from 412 files, a search filtered to a workspace's files returns the top 20 results in 10.3 ms (median, warm cache). An unfiltered search across the entire collection takes 27.1 ms, but Atarim never uses this in production.
The direct comparison, run by Atarim using MariaDB 11.8 and Qdrant 1.14 on the same machine, with default settings and real embeddings:
| Vectors | Qdrant filtered latency | MariaDB filtered latency | Recall (both) |
|---|---|---|---|
| 5,000 | 8.8 ms | 2.7 ms | 0.996 |
| 10,000 | 4.1 ms | 2.4 ms | 0.998 |
| 22,000 | 3.9 ms | 2.4 ms | 0.998 |
At 22,000 vectors with 16 concurrent clients, MariaDB served 1,292 queries per second versus 880 for Qdrant. On unfiltered searches, which use the HNSW index, the ratio reverses: Qdrant is faster and achieves 1.00 recall versus 0.89–0.94 for MariaDB. The dataset is small, so it doesn't yet show how both systems perform with larger volumes: a follow-up article promised by Atarim will use synthetic benchmarks at 50 times the current volume.
WordPress: the cost of vectorising a site
Atarim also offers clients database health checks, natural language queries, and actions performed by an LLM on WordPress data. The next step is semantically indexing content, but the additional storage varies considerably depending on what gets vectorised, measured on a synthetic database of around 1,200 posts, pages and products (95 MB):
| What gets vectorised | Storage added | Relative to site |
|---|---|---|
| All text, including revisions and comments | 552 MB | 5.9× |
| Published content only | 120 MB | 1.28× |
| Published, 768 dimensions | 68 MB | 0.72× |
| Published, 512 dimensions | 46 MB | 0.49× |
At 512 dimensions, recall drops to approximately 80% of that achieved with the 1,536 dimensions currently used by Atarim. Atarim's WordPress plugin has around 1,000 active installations in version 5.1.5, according to the official WordPress.org plugin directory.
Atarim and MariaDB will present the system in a live session on 3 December; registrations are on the MariaDB website.
Written with the help of artificial intelligence and checked by the editors (EU AI Act, art. 50).