Skip to content

ROADMAP.md

Latest commit

 

History

History
65 lines (53 loc) · 3.28 KB

ROADMAP.md

File metadata and controls

65 lines (53 loc) · 3.28 KB
Raw

Roadmap

A living document. ordvec is a small, paper-driven project; this captures direction and scope, not committed dates.

Vision

ordvec is a retrieval primitive, not a database. The goal is to be a universally embeddable building block for edge-deployed RAG and on-device / local AI retrieval — settings where compute, memory, and storage are tight and the hardware is heterogeneous (servers, ARM edge devices, browsers / WASM, and eventually mobile and embedded targets).

Everything distinctive about ordvec serves that end:

  • Training-free encoding — no codebook, no learned rotation, no fit step, and nothing to refit when the corpus drifts.
  • Zero system dependencies — pure Rust, no BLAS / LAPACK / ndarray / faer; no native library to install; links and cross-compiles cleanly.
  • Predictable footprint — exactly dim × bits / 8 bytes per document, known before you see any data.
  • Runtime-dispatched SIMD with a scalar fallback — the same retrieval core runs anywhere it compiles (AVX-512 / AVX2, NEON, WASM simd128, scalar).

Non-goals

ordvec will not grow into a standalone vector database. It deliberately does not pursue billion-scale navigable-graph ANN, distributed serving / sharding, a query language, or persistence and transaction machinery beyond its simple file formats. That is the territory of pgvector, Qdrant, Milvus, LanceDB, and full Graph-RAG stacks. ordvec stays the substrate those systems — and bespoke edge pipelines — compose, not a competitor to them.

Direction

The throughline is "be a good neighbour": ordvec should embed natively into more hosts rather than forcing callers to adapt to it.

  • Distribution baseline. ordvec now has coordinated crates.io / PyPI release machinery with SLSA build provenance, SBOM artifacts, registry trusted publishing, and post-publish byte checks. Ongoing work here is platform breadth, not the first publish path.
  • Cross-stack embedding via C and Go. The repo now carries a base publish=false C ABI (ordvec-ffi) plus a thin Go wrapper (ordvec-go) for loading persisted RankQuant / Bitmap indexes and running synchronous search. Next ABI work is the v2 builder/add surface, richer batched APIs, and compatibility policy.
  • Adapters. Thin integration layers for host retrieval / RAG systems. ordvec factored out of turbovec; natural next targets are mainstream RAG frameworks (via the Python binding) and Rust-native edge Graph-RAG stacks such as EdgeQuake, where ordvec could back the vector-search layer in a Postgres-less / edge deployment — complementing, not replacing, the graph layer.
  • Broader platform reach. musllinux wheels (Alpine / minimal containers); continued aarch64 and WASM coverage.
  • Toward 1.0. An API-stability pass and a documented compatibility policy.
  • Supply-chain posture. OpenSSF Best Practices passing is earned; silver is the next target (governance, this roadmap, coverage reporting).

Out of scope here

Benchmarks and the intellectual framing live in the OrdVec / RankQuant paper and the companion ordvec-formalization Lean repository; this roadmap concerns the crate as an embeddable building block.