Architecture

BudgetVec is built as a Rust → WASM application running on Cloudflare Workers, with R2 for durable storage and KV for caching and control plane data.

System Diagram

┌──────────────┐ │ Client │ │ (SDK / HTTP) │ └──────┬───────┘ │ HTTPS ▼ ┌──────────────────────────┐ │ Cloudflare Worker │ │ (Rust → WASM, 691 KB) │ │ │ │ ┌─────────────────────┐ │ │ │ Router / Auth │ │ │ │ Tenant isolation │ │ │ └─────────┬───────────┘ │ │ │ │ │ ┌─────────▼───────────┐ │ │ │ Ingestion Pipeline │ │ │ │ Validate → Encode │ │ │ │ → WAL batch │ │ │ └─────────┬───────────┘ │ │ │ │ │ ┌─────────▼───────────┐ │ │ │ Query Engine │ │ │ │ ANN (SPTAG/PQ ADC) │ │ │ │ BM25 (Thai + EN) │ │ │ └─────────┬───────────┘ │ │ │ │ │ ┌─────────▼───────────┐ │ │ │ Pufferfish Cache │ │ │ │ L1: Worker memory │ │ │ │ L2: KV │ │ │ │ L3: R2 │ │ │ └─────────────────────┘ │ └──────────────────────────┘ │ ▼ ┌──────────────┐ ┌──────────────┐ │ R2 Storage │ │ KV Store │ │ $0.015/GB │ │ Control + │ │ Vectors/WAL │ │ Cache │

└──────────────┘ └──────────────┘

Crate Structure

BudgetVec is organized as a Rust workspace with 6 crates:

Crate	Purpose
`sptag-sys`	C++ FFI bindings for SPTAG indexing
`sptag`	Safe Rust wrapper around SPTAG
`core`	Codecs (PQ, RaBitQ), storage, text processing, BM25
`worker`	Cloudflare Worker entry point, routes, platform code
`build-cli`	Offline index builder CLI tool
`tests/http-integration`	HTTP integration test suite

Compression Pipeline

Vectors go through a two-stage compression:

1. RaBitQ — 1-bit quantization reduces each float32 dimension to 1 bit, giving 32× compression for distance approximation

2. Product Quantization (PQ) — Subdivides vectors into M subspaces, quantizes each to K centroids, enabling ADC (Asymmetric Distance Computation) for fast approximate search

Multi-Tenant Isolation

Each tenant has:

Unique API key (FNV-1a hashed for lookup in KV)
Isolated R2 prefix: {base_namespace}/{namespace}/...
Per-tenant KV cache entries
Configurable vector dimensions, plan limits, and default region

Regions

BudgetVec supports 6 deployment regions, each with dedicated R2 buckets and KV namespaces:

Region	Location
WNAM	US West
ENAM	US East
WEUR	Western Europe
EEUR	Eastern Europe
APAC	Asia-Pacific
OC	Oceania