Skerpa test bench

Ingestion on skerpa's high tier: run the pipeline as it ships, with Jev doing the dedup, or fully Jev — and compare them on entities, quality, speed and cost.

Before you start

checking…

1 · Your file

Document to benchmark

pdf, docx, pptx, md, txt, html, xlsx. This is the one that gets timed.

A related document to load first (optional but recommended)

The dedup judge only runs when the project already holds entities. This document is ingested and merged first, untimed, so the judge has something to compare against. Leave empty and the report will say the judge was not exercised.

2 · What to run

Tick one or more. Ticking option 1 as well makes it the yardstick for the others in this same run; leave it unticked and they use your last saved option 1 for this file.

3 · Options

Document type (skerpa's uploader hint)

Language tag

TypeSafe's model is English-first. For other languages, read the agreement rate before the speed.

What is the conversation about? (optional)

The yardstick for the "off topic" judgment. Left empty it is inferred from the transcript's title or its middle — never its opening, because calls start with greetings and everything substantive would then look off topic.

Repeats

LLM latency varies run to run. 3 gives medians; each repeat costs time and money.

Skerpa test bench

Before you start

1 · Your file

2 · What to run

3 · Options

Running

Decisions, live

Result

Earlier runs