From a3bbee4d5179af7699e60a5c5053f0014f791312 Mon Sep 17 00:00:00 2001 From: Nelson Spence Date: Tue, 16 Jun 2026 14:32:16 -0500 Subject: [PATCH 1/2] docs: add "What is ordinal retrieval" section to README Add a top-level conceptual frame above "## Benchmark at a glance" so the README reads define-the-idea -> prove-it. It previously jumped straight from "what the crate is" to the benchmark without ever stating the idea; this states what ordinal retrieval is (the index operates on order/sign quotients of embedding coordinates rather than metric magnitudes; training-free compression, quotient-aware candidate generation, asymmetric rerank) right before the benchmark evidence. The "matches dense qrel quality at much smaller storage" line restates the existing in-repo BEIR result (the hero claim immediately below it), not a new or paper number. Signed-off-by: Nelson Spence --- README.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/README.md b/README.md index 2afed35c..c43e5e7b 100644 --- a/README.md +++ b/README.md @@ -16,6 +16,20 @@ Training-free ordinal & sign quantization for vector retrieval. that quantizes the **ordinal (rank) and sign structure** of an embedding — no codebook, no learned rotation, no graph to build. +## What is ordinal retrieval + +Ordinal retrieval is a retrieval family where the index operates on order/sign +quotients of embedding coordinates rather than metric magnitudes, with +training-free compression, quotient-aware candidate generation, and asymmetric +reranking against the original query signal. + +Dense retrieval assumes magnitude geometry is the primitive. Ordinal retrieval +asks what survives when embeddings are quotiented down to coordinate order, +signs, and relative salience. Surprisingly, enough retrieval signal survives to +match dense qrel quality at much smaller storage, and those quotients support +quotient-aware candidate-generation rules that are training-free, +append-friendly, and graph-optional. + ## Benchmark at a glance > **ordvec matches dense retrieval quality within BEIR qrel noise at 8–16× smaller From ab54df39caaa6c8f5ccdc636d617d475541ace80 Mon Sep 17 00:00:00 2001 From: Nelson Spence Date: Tue, 16 Jun 2026 15:13:48 -0500 Subject: [PATCH 2/2] docs: add question mark to "What is ordinal retrieval?" heading gemini review nit (PR #245): the heading is phrased as a question, so the trailing ? is grammatically correct. Signed-off-by: Nelson Spence --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index c43e5e7b..98f6ed00 100644 --- a/README.md +++ b/README.md @@ -16,7 +16,7 @@ Training-free ordinal & sign quantization for vector retrieval. that quantizes the **ordinal (rank) and sign structure** of an embedding — no codebook, no learned rotation, no graph to build. -## What is ordinal retrieval +## What is ordinal retrieval? Ordinal retrieval is a retrieval family where the index operates on order/sign quotients of embedding coordinates rather than metric magnitudes, with