VQS

VECTOR QUALITY STANDARDS A score for the corpus itself, before anyone asks it a question.

What the score is

VQS reads the chunks you stored and scores the corpus on its own terms. No question, no judge, no answer key. Just the material: how the chunks sit in vector space, how well each one is written, and where they disagree with each other.

What it tells you

Where retrieval is going to struggle, before your users find it for you. A per-document score names the documents dragging the corpus down. The quality map shows you whether they cluster or scatter. The atlas in the dashboard draws the whole corpus, colours it by quality, and arcs the contradictions in red between the chunks that disagree.

What it is not

VQS is a function of your corpus. Whether the corpus answers a particular question is a function of the corpus, the question, and the model reading it. Those are different quantities and we do not pretend otherwise. A contracts corpus can be in excellent shape and still have nothing at all to say about last quarter's pricing. VQS scores what the corpus is, never what it can answer.

SCORED ON THE ENGINE THAT SERVES YOUR QUERIES WHAT MOVES THE SCORE → WHAT IT DOES NOT CLAIM → > knowledge is precious

Geometry

Query-free measures over the vectors themselves: cohesion, spread, margin, hubness, redundancy, integrity, and split penalty. This is where chunk-splitting damage and near-duplicate pileups surface, the two failures that quietly cost you the most recall.

Writing quality

Per-chunk text measures: lexical diversity, readability, information density, punctuation density. A chunk that reads like a fragment of a table of contents scores like one, because that is what it will retrieve like.

What to do about a low score

Four moves, all of them yours. Split or rewrite the chunks flagged as incoherent. Drop the near-duplicates so they stop competing with each other. Give every fact that varies by version, date, or tier its own key in the text, so the chunk stands on its own instead of reading as a contradiction. Then add material where the map is thin.

Which layer your problem is in

This is the part worth the read. A low score means better retrieval will not save you: the material is the bottleneck, and fixing it comes first. A high score with answers still going missing is a retrieval problem, and that is where Max retrieval at $45 a month earns its keep. Buying Max to paper over a weak corpus is money lit on fire, and we would rather say so than sell it to you.

Scored on the real engine

VQS is not a heuristic run over your text in a notebook. Your corpus is embedded by the same engine that answers your queries, on the same owned NVIDIA DGX hardware, so the geometry you are scored on is the geometry retrieval actually sees.

Live in the dashboard

Open Answers, then Quality. Per-document scores, the 2D quality map, and a 3D atlas of the corpus: structure as you built it, quality as colour, contradictions drawn as red arcs between the chunks that cannot both be right.

What it does not claim

VQS is a relative signal, not an absolute certified scale. It compares corpora, and it compares your corpus to itself over time. It does not certify a corpus as good in some universal sense, and it will never promise that a given question has an answer in your documents. The set of questions people can ask is open-ended, so a corpus in perfect shape still has edges.

Score a corpus

Upload documents, then open Answers and Quality. Turbo embeddings are free forever and your first 5,000 documents are included, so the first score costs nothing.

Create free account 60 SECONDS · NO CARD · NO SALES CALL

Numbers, with receipts: 87ms is P50 end to end on owned NVIDIA DGX hardware, no shared queue. English MTEB v2, Borda 75.99, rank #1, independently verified on the public leaderboard. Those figures rank the embedding engine VQS runs on, not VQS itself. VQS carries no public benchmark and is published here as a relative signal only. The readout above is an illustration, not a customer corpus. Methodology in the docs.