Sprint 1 demo script — Sep 18

The script for Friday’s demo of Sprint 1, against the production host and nothing else. Six steps, under five minutes. Every step is a page the host serves or a request the API already answers; nothing is staged, and no step needs a terminal.

Paths below are the host’s own. <host> is the production hostname; the app lives at the root and the documentation under /docs/, ADR-0056.

Before the room

Open https://<host>/en/ and encode each of the three samples the script uses, so the demo runs off the host’s encode cache rather than off a cold start. That is the only preparation, and it changes nothing a visitor would not get on their second click.

1. The advantage on one page — 30s

Open https://<host>/. It redirects to /en/, the sales page: the headline, the two buttons and the six steps, each with one sentence and a link to the run that proves it.

Say the headline, then click the language toggle once to show /ja/ and come back.

2. Every page understood, and the best parser per page — 90s

Click the proof under step 3, “The best parser per page, and the reason”. It opens the playground on ?sample=mhlw-notice&run=encode and encodes a 2021 MHLW administrative notice.

• Page 1: the plan chose native-extraction, the reason is the text layer, and the fallbacks it kept are listed under it.
• Page the pager to 5, the scanned attachment: the plan chose generic-ocr, the reason is that there is no text layer, and the line says whether it was read as planned.

One document, two parsers, decided before anything expensive ran. This is the scenario the deploy itself asserts (@deployed, tests/bdd/features/encode.feature).

3. One structured result — 60s

Stay on the run and move across the views:

• Entities — the spans marked in the page’s own text.
• Tags — the facets the classifiers answered and the tags the enrichers wrote.
• Chart — open ?sample=process-chart&run=encode instead: slide 2’s process chart, its symbols and connections, read as a graph (ADR-0043). Then click Draw.io file and open what downloads: the same graph, in the editor the customer’s engineers already use. Say that the parser’s distribution writes the file (ADR-0058), which is what “your formats, from your examples” means before a recipe exists.

4. Ready for search — 45s

On the same run, open Search and search a phrase from the document. A hit names its page and its box, and says which lane found it: the words the chunk carries, the vector POST /v1/embed returned, or both.

Say the scope out loud, because the page says it too: the index is the documents this browser encoded, held in this browser, never sent anywhere.

5. The same document again — 15s

Encode the same sample a second time. It comes back immediately, labelled “served from cache, computed at …”. Say what it is: a convenience of this demo host, not a product feature — nothing in indx serve caches a result.

6. The contract underneath — 30s

Open /docs/ for the documentation, and /scalar for the API reference the screens are generated from. Then the deployment screen at /en/deploy/, which reads the live capability snapshot: this is the same service, in one container, in the customer’s own cloud.

What to say if asked

• Accuracy? Not yet. The labels behind the accuracy figures are under review, and the benchmarks page publishes only the three that depend on no label.
• Other clouds? Terraform for AWS, Azure and GCP; AWS is deployed and serving. The other two are what Sprint 2 reports on.
• Our own output format? Step 5 of the sales page: configuration and examples, not engineering. This week’s proof is the Draw.io file; the few-shot library is next.