Same word. ⅛ the bytes. The brand is the product demo, said out loud.
That choice was made when embeddings were 128-dim and corpora held 100,000 records. It hasn't been revisited. A 1,536-dim OpenAI vector is now 6,144 bytes on disk, in WAL replication, in object-storage tiers, and in every byte of the working set the buffer cache has to hold. That's the bill you pay every month. Storage and IO together drive 55–80% of vector-database cost.
A 100-million-vector retrieval corpus at typical embedding dimensions: 614 GB of raw embedding bytes — on disk, on every replica, in every backup, in every snapshot. Across managed services and self-hosted clusters, that footprint scales linearly with both your data and your dimension count. CRBRL stores the same vectors in 77 GB. Same retrieval quality. Migrate in a weekend.
CRBRL ships as a standalone vector database for greenfield deployments, and as a Postgres extension for teams already running on Postgres. Same engine, same retrieval quality — different operational shape.
A vector database for greenfield deployments. Compatible with the Chroma API, so applications already using Chroma migrate without code changes. Same database from development through production scale.
crbrl-pgAn extension for your existing Postgres database. The crucial difference: search runs directly on the compressed index — no decompression in the query path. That's what makes it efficient at scale, and why bolt-on compression in other engines never matched it.
Compression mathematics that work on the first vector you load. No training. No calibration. Peer-reviewed and proven near information-theoretic optimal.
Recent data stays at higher fidelity. Older data compresses further as it ages. Automatic, in the background. No operational work.
Semantic search, full-text, and hybrid retrieval — all in one query interface. All running directly on compressed data.
OpenAI, Gemini, Cohere, Mistral, and more — built in. Switch providers without rebuilding your pipeline. Use whichever fits the workload.
Authentication, audit logs, multi-tenancy, observability, and snapshots. Built in from day one, not bolted on when customers ask.
Three tiers, automatically managed. Recent data stays at higher fidelity. Older data compresses further as it ages. No operational work — the system shifts records as their access pattern changes.
AI infrastructure is being rebuilt around retrieval. Vector databases are the fastest-growing line item in that rebuild, and storage is the fastest-growing line item inside vector databases. Compression is the lever no incumbent has pulled yet — because their engines weren't designed for it.
From $91B in 2026 at 23% CAGR. The data-substrate layer (vector retrieval, graph, agent memory, cache) goes from $12B to $27B in the same window — 31% CAGR, faster than the broader market. Source: Precedence Research, Jan 2026.
Embedding dimensions doubled in 24 months — 768 → 1,536 → 3,072. Corpora grew super-linearly. The bytes scale by both axes; the bills compound monthly. Hot-cache hit rates fall as corpora grow. None of this is a feature.
Every major vector database stores embeddings as 32-bit floats — the default since the category began. That choice was made when embeddings were 128-dimensional. Peer-reviewed mathematics now compresses the same vectors 7–8× with no retraining and no recall loss. The first lever in the category that doesn't trade quality for cost.
Move the buttons. The math runs on your inputs against the same compression ratio TurboQuant produces in the codebase. No registration. No "talk to sales." Numbers update live.
Standard benchmarks. Same retrieval quality at every compression level — pick the trade-off that fits your workload.
If you're on Chroma, point your existing application at CRBRL — the API is compatible. If you're on Postgres, add the extension to the same database. Five minutes to first query, either path.
We're working with a handful of teams whose retrieval bills aren't aging well. Tell us a bit about your stack — we'll be in touch.