Fieldnote-Echo

Fieldnote-Echo / ordvec

Public

Training-free ordinal & sign quantization for compressed nearest-neighbour retrieval over high-dimensional embeddings. Pure Rust, zero system dependencies.

ordvec.dev ann approximate-nearest-neighbor embeddings information-retrieval nearest-neighbor
10
3
100% credibility
Found May 27, 2026 at 10 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Rust
AI Summary

ordvec is a training-free library for compressed nearest-neighbor search over high-dimensional embeddings. It compresses vectors based on the rank order of their coordinates (rather than learning patterns from data) and supports fast similarity search with adjustable precision. The library provides four index types: Rank for full-precision ranking, RankQuant for compressed bucketed ranks, Bitmap for fast candidate filtering, and SignBitmap for sign-based retrieval. It includes SIMD acceleration for x86_64, ARM, and WebAssembly targets, and ships with Python bindings for easy integration.

How It Works

1
💭 You have embeddings to search

You've built an AI application that creates vector representations of documents, images, or text — and now you need to find similar items quickly.

2
📦 You add the search library

You install ordvec as a dependency in your project. It works right away with no extra setup or training required.

3
🔢 You feed in your vectors

You add your collection of embeddings to the index. The library compresses each one based on the ranking order of its values.

4
⚡ Everything compresses instantly

Without any training or fitting, your vectors are compressed and ready to search. The library knows exactly how much space each one takes.

5
🔍 You ask to find similar items

You enter a query vector and ask for the top 10 most similar items from your collection.

6
You can choose how to search
🎯
Direct search

Search directly through your compressed vectors for the best matches

🔎
Two-stage search

Quickly filter candidates with a bitmap, then rerank the best ones with full precision

✅ You get your results

The library returns the most similar items ranked by relevance, using the ordinal structure of your vectors.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 10 to 10 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is ordvec?

ordvec is a Rust library for compressed nearest-neighbor search over high-dimensional embeddings. It quantizes the rank order of vector coordinates (rather than their values) to enable fast retrieval without any training step. The library ships with four index types: full-precision ranks, bucketed ranks at 1/2/4 bits per coordinate, top-bucket bitmaps, and sign bitmaps. SIMD acceleration handles the scanning workload across AVX-512, AVX2, NEON, and WebAssembly targets. Python bindings are available via PyPI.

Why is it gaining traction?

Most vector search libraries require fitting a codebook to your data before you can index anything. ordvec skips this entirely -- encoding is per-vector and works on the very first document with no prior data. The memory footprint is predictable before you see any data: at 2 bits per coordinate and dimension 1024, you get exactly 256 bytes per document. The two-stage retrieval pipeline (bitmap prefilter feeding an exact rerank) ships as built-in primitives rather than requiring you to wire together separate components. The theoretical backing is unusual -- bitmap selectivity has a closed-form hypergeometric null, and key invariants are machine-verified in Lean 4.

Who should use this?

Backend engineers building retrieval systems who want a lightweight, dependency-free substrate. Researchers working with embedding pipelines who need reproducible, deterministic quality numbers. Teams serving vector search on constrained targets (aarch64, wasm32) where system dependencies like BLAS are unavailable. Anyone evaluating whether ordinal structure alone is sufficient for their recall requirements -- the synthetic benchmark is reproducible and self-contained.

Verdict

At version 0.2.0 with 10 stars, ordvec is early-stage and the broader real-corpus evaluation is still in progress. The 1.0% credibility score reflects that maturity. The documentation is thorough and the reproducible benchmark is a genuine differentiator, but this is a library for builders comfortable with experimental tooling. Worth watching if you need training-free compressed retrieval or are researching ordinal quantization as a signal.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.