Op2ra → the documents surface → browser audit
Nullius · free, and nothing is uploaded

Audit a retrieval test set, here, now.

This runs the same calibrated engine a paid engagement runs — op2ra 0.4.0, installed into a Python runtime inside this tab. Your corpus and your queries are read by JavaScript on your machine and handed to that runtime in memory. They are never sent anywhere, including to us.

What actually happens when you press run

Open your browser's network tab and check — this is the whole list, and it does not change based on what you select.

  • Pyodide 0.26.4 is downloaded from jsdelivr — a CPython build compiled to WebAssembly, about 12 MB, cached by your browser after the first run. This is a third party, and it is the only one.
  • numpy is downloaded from the same place — the one runtime dependency this project has.
  • The op2ra wheel is downloaded from op2ra.com — op2ra-0.4.0-py3-none-any.whl, 76 KB, the same artifact pip install would give you. Not a JavaScript reimplementation: a reimplementation would be a second instrument with none of the calibration behind it.
  • Nothing else. There is no analytics script, no upload endpoint, and no request carrying a single byte of your files. There is nowhere for them to go — this site is static hosting with no server to receive them.

The tradeoff is honest and worth stating: because it runs in your tab, a large corpus is limited by your own memory, and the first run pays a one-off download an engagement does not.

Run it

Bring your own, or start with the samples.

The samples are two evaluation sets over one corpus of twelve documents. They ask the same twelve questions and are the same size. The only difference is that one was written by opening the documents and lifting their wording, and the other was written from what somebody would actually type. Run both.

A folder of documents, or a JSONL file of {"id": ..., "text": ...}. Same shape op2ra documents audit takes on disk.
Selecting a folder not working?
JSONL, one per line: {"query": ..., "relevant": ["doc id", ...]}. gold works as an alias for relevant.

What the browser version cannot do

The section a competent buyer reads first, and it applies to this page too.

  • It audits the test set you already have. If the test set is the problem, that is the finding — but it cannot write you a better one, and building one that is not contaminated is most of the work in an engagement.
  • It reads what you hand it. It cannot see the queries you did not include, or tell you whether this corpus is the one your users actually search.
  • It does not score your retriever. Nullius audits the ruler. A retrieval number is only meaningful once the set producing it survives this, and that ordering is the entire product.
  • It is one surface. The findings that explain a number usually live next to it — in how the experiment was powered, or how often it was checked — and those are different audits.

Engagements

Fixed scope, fixed price, and a report you can argue with.

Free

Browser audit

The documents surface, running in your browser — the real engine, installed into the page. Nothing is uploaded, because there is no server to upload it to. Or send one artifact and we will look at it: you get the finding either way, including if the finding is that nothing is wrong.

Fixed

One surface, two to three weeks

One product run against your artifacts, with a written report: measures, findings by severity, the resolution floor, and the limits. The price is fixed before the work starts and quoted from the size of your company, not from how the conversation goes.

Quote

Multiple surfaces

Where it gets interesting — the findings on one surface routinely explain the numbers on another.

Found something?

Send the report, or the artifact behind it. The first look costs nothing and the finding is yours either way — including if the finding is that your test set is fine.

rishabh@op2ra.com