CRBRL /ˈsɛr.ɪ.brəl/ · COMPRESSION

COMPRESSION-NATIVE · DISK-PERSISTENT · VECTOR + GRAPH

Store 8× more on disk. Retrieve at the same fidelity. Pay a fraction.

CRBRL is a compression-native, disk-persistent vector database. Its codec layer is peer-reviewed (arXiv:2504.19874, ICLR 2026) — so the claims below are numbers, not adjectives.

FEWER BYTES
PER VECTOR
≈0.00
COSINE FIDELITY
AFTER COMPRESSION
0 B
PER 1,536-DIM VECTOR
VS 6,144 B FLOAT32
0
TRAINING PASSES
ZERO-TRAINING CODEC
CH 01THE PROBLEM, MEASURED

Embeddings grew into an economics problem.

One 1,536-dimension embedding stored as float32 is 6,144 bytes. At 100 M vectors that is 614 GB — and RAM-first engines ask you to keep it hot to stay fast.

The industry answer has been bigger clusters. The math answer is fewer bits.

1,536
DIMENSIONS · TYPICAL
EMBEDDING (E.G. OPENAI)
6,144 B
PER VECTOR · FLOAT32
BEFORE COMPRESSION
614 GB
PER 100 M VECTORS
HELD AT FULL PRECISION
CH 02THE CODEC · ZERO-TRAINING COMPRESSION

TurboQuant: 8× fewer bytes, ≈0.98 cosine fidelity, no training step.

TurboQuant is data-oblivious: no codebook, no training pass, no drift as your corpus changes. A random rotation plus Lloyd-Max quantization (PolarQuant), corrected by a 1-bit QJL inner-product estimator, takes a 1,536-dim vector from 6,144 B to 768 B at ≈0.98 cosine recall.

QUANTIZATION BENCH — LIVE 1,536-DIM VECTOR · 2-D PROJECTION
BYTES / VECTOR
768 B
COSINE FIDELITY
≈0.98
100 M VECTORS ON DISK
76.8 GB

Point cloud is a 2-D projection for illustration. Fidelity figures for the 4-bit operating point are from arXiv:2504.19874; intermediate points interpolated.

TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate
Zandieh · Daliri · Hadian · Mirrokni — Google Research + NYU
arXiv:2504.19874 · published at ICLR 2026
CH 03CODEC CHOICE · PER MEMORY CLASS

Two shipped codecs. Pick per memory class, not per vendor roadmap.

Compression is CRBRL's substrate, not a mode you switch on. The codec layer is selectable — benchmark a corpus against both, assign the better fit to each memory class.

DEFAULT

TurboQuant

Data-oblivious. No training, no codebook, no drift. Maximum storage reduction: 8× fewer bytes per vector at ≈0.98 cosine fidelity.

arXiv:2504.19874 · ICLR 2026
BEST FOR: max density, changing corpora
SELECTABLE

RaBitQ

Randomized bit-quantization with an unbiased distance estimator and a provable error bound — for workloads where accuracy must be guaranteed, not just demonstrated.

Gao & Long · SIGMOD 2024
BEST FOR: high-dim ANN with hard accuracy floors
CH 05ONE ENGINE · VECTOR + GRAPH

Similarity says what's related. The graph says how. One engine answers both.

CRBRL ships vector search and relationship intelligence in the same engine — one query, one store, one operational surface. The prevailing alternative is stitching a vector database to a separate graph database and reconciling two systems of record.

VECTOR MODE

Semantic retrieval

Compressed-domain ANN over HNSW, hybrid lexical + semantic ranking with BM25 + RRF. Built for RAG and agent memory read paths.

SHIPPED · v0.1.0
GRAPH MODE

Relationship intelligence

Entities and edges live beside the vectors they describe. Traverse how documents, people, and events connect — without exporting to a second database.

SHIPPED · v0.1.0
CH 06MEMORY TIERS · MULTIPLICATIVE

Tiering and compression multiply. Others make you choose.

Hot, warm, and cold classes each carry their own codec setting. Cost engineering that tiers full-precision vectors saves on where bytes sit; CRBRL first shrinks the bytes, then tiers them.

HOT

Active working set — lowest-latency reads, codec tuned for fidelity.

NVME / MEMORY-MAPPED
WARM

Recent but not resident — compressed-domain reads straight off disk.

LOCAL DISK
COLD

Long-tail memory — maximum-density codec settings, still queryable.

DENSE STORAGE

8× compression × tiered placement = compounding, not either/or.

System-level cost outcomes are workload-dependent — the 8× / ≈0.98 codec figures are the peer-reviewed constant; what they compound into depends on your read mix and corpus shape.

CH 07DATASHEET · WHAT SHIPS TODAY

Two products. Both shipped, both under one brand.

Everything on this sheet is v0.1.0 capability — no roadmap items listed. Design partners are in flight.

CRBRL · STANDALONE

SINGLE DOCKER IMAGE · .NET 8

API SURFACEREST v1/v2 (Chroma-compatible), gRPC, native MCP server
RETRIEVALHNSW ANN + BM25 lexical + reciprocal-rank fusion
CODECSTurboQuant (default) · RaBitQ — selectable per memory class
GRAPHEntities + edges beside vectors; one query surface
EMBEDDINGS9 providers supported out of the box
OPERATIONSMulti-tenant RBAC · snapshots · distributed sharding
PERSISTENCEDisk-first; compressed-domain reads off disk

crbrl-pg · POSTGRES EXTENSION

RUST · PGRX

WHAT IT ISTurboQuant vector types + HNSW operator classes inside Postgres
MIGRATIONpgvector workloads move without a database swap
WHYKeep transactions, backups, and tooling where they already live
FOOTPRINT8× fewer bytes per vector inside your existing tables
STATUSShipped · design partners in flight
CH 08DROP-IN · NO REWRITE

Arrive without a migration project. Leave nothing behind.

CRBRL speaks the interfaces you already use: Chroma-compatible REST for existing clients, a native MCP server for agents, and a Postgres extension for pgvector estates.

STANDALONEDOCKER
$ docker run -p 8000:8000 crbrl/crbrl:0.1.0

# Chroma-compatible REST — existing clients just point here
$ curl localhost:8000/api/v2/heartbeat
{"nanosecond heartbeat": 1753077600000000000}

# agents connect over native MCP — no adapter layer
POSTGREScrbrl-pg
postgres=# CREATE EXTENSION crbrl;
CREATE EXTENSION

-- pgvector migration without a database swap:
-- TurboQuant types + HNSW opclasses, in-place
-- 6,144 B → 768 B per 1,536-dim vector
CH 09CONTACT · DESIGN PARTNERS

If your vector bill is a line item you flinch at, send us the shape of it.

Tell us dimensions, corpus size, and read mix — we'll reply with what the codec math says it becomes on CRBRL. Design-partner slots are limited by engineering attention, not sales capacity.

PRESS ~ TO TUNE SCROLL