New direction — concept for review · static preview, no backend
CRBRL · Rev 2026.07 · Precog Labs
Datasheet · Compression-native vector database
CRBRL

Four bits per dimension. 8× fewer bytes on disk. ≈0.98 cosine fidelity retained.

Codec
TurboQuant
Encoding
4 bits / dim
Fidelity
0.98 cosine
Proof
arXiv:2504.19874
§0 Abstract

What this is, in one paragraph

CRBRL is a compression-native, disk-persistent vector database built around TurboQuant — a two-stage, training-free quantizer that needs no codebook and no per-corpus calibration. Stage one (PolarQuant) applies a random orthogonal rotation so every input distribution concentrates on the same fixed Beta curve, then quantizes each coordinate with a scalar quantizer fit once at codec init. Stage two (QJL) retains one sign bit per dimension for unbiased inner-product estimation, enabling similarity search directly on the compressed codes — no decompression step. At the default configuration (b = 3) the total cost is 4 bits per dimension: one eighth of the float32 baseline, at ≈0.98 cosine fidelity. CRBRL ships as a standalone server (Docker, Chroma-compatible REST, native gRPC, MCP) and as crbrl-pg, a Postgres extension that adds TurboQuant column types to an existing pgvector deployment with no data migration.

[1] Zandieh, Daliri, Hadian, Mirrokni — ICLR 2026 · arXiv:2504.19874
§1 Fig. 1

Storage cost per vector, by encoding

32 bit
float32 baseline
16 bit
half precision
8 bit
generic PQ
4 bit
CRBRL / TurboQuant
Fig. 1 — bits stored per vector dimension at default settings (b = 3). Lower is better; recall is not traded away at the last step the way generic 1-bit binary quantization does. Source: arXiv:2504.19874, §5.1.
§2 Method

How a vector becomes 4 bits

Input
Raw float32 vector, any distribution
Stage 1 · PolarQuant
Random orthogonal rotation, then a fixed scalar quantizer per coordinate
Stage 2 · QJL
A second JL-style projection, sign bit retained per dimension
Output
Compressed code, 4 bits / dim, disk-persisted
Query
Similarity computed on the compressed code directly — no decompression
No training data, no codebook, no per-corpus calibration — the same quantizer works on any input because the rotation in stage one washes out the distribution before quantization happens. That's what makes it a database primitive rather than a tuning exercise: CREATE EXTENSION, swap the column type, done.
§3 Results

Monthly cost, 100M vectors, self-hosted on AWS

MetricPinecone ServerlessChroma self-hostedCRBRL self-hosted
Storage footprint—3,072 GB384 GB
Compute—3× r6i.8xlarge2× r6i.4xlarge (smaller working set)
EBS spend / mo—$246$31
Monthly total~$6,000–7,000~$6,246~$2,031
Model uses on-demand AWS pricing, excludes reserved/committed discounts and operational labor. CRBRL applies the 8× compression factor uniformly to disk, WAL, replica, and snapshot. Annualized savings at this scale exceed $50,000 per workload. Full methodology and assumptions available on request. — Source: CRBRL TCO model, §7.
§4 Deploy

Three ways to run it

CRBRL-OSS

Self-hosted

Open source under a permissive license. Docker image or crbrl-pg as a Postgres extension — column-type swap, no migration.

Get the image
CRBRL-ENT

Enterprise / BYOC

Self-hosted or bring-your-own-cloud, with support, audit-grade compliance evidence, and a reference architecture for regulated deployments.

Talk to us
CRBRL-CLOUD

Managed cloud

Usage-based, $/GB-month and $/operation. Private beta targeted late 2026 — the compression-native cost basis carries straight through.

Join the beta
§A Appendix

Standing questions

A.1Is TurboQuant lossy, and does that matter for recall?＋

Yes, it's lossy — every quantization scheme is. The difference is that TurboQuant has proven error bounds at every bit rate, so the loss is a known, chosen number (≈0.98 cosine fidelity at the default setting) rather than an emergent property you discover in production.

A.2How is this different from product quantization or binary quantization?＋

Generic PQ needs a codebook trained per corpus. Generic 1-bit binary quantization is fast but trades away meaningful recall. TurboQuant needs no training, works on any input distribution, and gives an unbiased inner-product estimator at 4 bits/dim rather than 1.

A.3Does crbrl-pg require a database migration?＋

No. It's a Rust extension on pgrx that adds TurboQuant column types and HNSW operator classes alongside pgvector. Adoption is CREATE EXTENSION plus a column-type swap.

A.4What's the licensing model?＋

The codec, engine, and Postgres extension are open source under a permissive license. Revenue comes from managed cloud, enterprise support/BYOC contracts, and a partner channel — the same open-core pattern HashiCorp, MongoDB, and Elastic built on.

A.5When is managed cloud available?＋

Private beta is targeted for late 2026. Self-hosted and enterprise/BYOC are available now.

§5 Cite this

For the readers who came for the paper, not the pitch

@inproceedings{turboquant2026, title = {TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate}, author = {Zandieh, Amir and Daliri, Majid and Hadian, Amirhosein and Mirrokni, Vahab}, booktitle = {International Conference on Learning Representations (ICLR)}, year = {2026}, eprint = {2504.19874}, archivePrefix = {arXiv} }