The no-code builder¶

A local wizard that turns a folder of recordings and CSVs into a working dashboard, without a terminal beyond the one command that starts it. It is the path for the researchers DIMS is for, who are domain experts and not necessarily programmers.

pip install -e './dims[builder]'
dims-builder

Your browser opens on step 1. Flags, ports and launchers are on the dims-builder reference.

It writes an ordinary study. Nothing about the result is wizard-specific: the skeleton comes from the same dims_case.core the dims-case command uses, so a study you build here and one you create on the command line are the same thing, and either can be continued with the other.

The seven steps¶

1 · Study¶

New or existing, where it goes, and public / private — the visibility decision, which you answer once and which is not inferred. Then the study's title, subtitle, authors, contacts and defaultWindowSize.

2 · Sessions & files¶

Drop files in. Each is given a role inferred from its extension, a session id and — for a time series — a measure name; you correct any of that in the table. This is where videoIDs and dataTypes come from.

"Load the example study" builds a complete synthetic study called ConvoConnect-Mini — two dyads, hand and rtpjSync measures, transcripts, an .eaf and generated video — so you can see the whole path before committing your own data. It is generated on demand, not shipped, and it is byte-stable: generating it twice gives identical files.

Studies filmed from several angles configure perspectives here.

3 · Align¶

The step that earns its place. Per session, trim the video and zero-pad the time series until the two cover the same span, with live extent bars showing where they currently disagree.

This is the one requirement DIMS cannot check for you and cannot recover from: each signal's time axis must map onto the video clock. A misalignment is displayed faithfully as a misalignment, and it is easily mistaken for meaningful dynamics.

It records specs, not edits — your original files are not modified.

4 · Tabs & analyses¶

Which analyses to run, and on what.

RQA, cross-RQA and cross-wavelet get chip pickers for measures and pairs, plus the advanced settings that become the analysis block.
The network diagram is drawn here: place measures on a body, drag between two of them to ask for that coupling. It is the same figure the dashboard draws, from the same geometry file, and the pairs you draw are the same list as the cross-wavelet chips above — either view edits it and both redraw.
ELAN is a toggle.

Switching the network on forces cross-wavelet on, because the network is a view of it. It also sets mcCount to 100 — but only when you have not set one; an existing non-zero value is left alone. The network's coherence mode is the only thing that reads the coherence null and is meaningless without it; its shared power mode does not need one.

5 · Build¶

Validation problems first — a series with no CSV and a duplicated series are errors; an unknown body part and an undefined group are warnings — then the assembled config.json for review, then write.

The config is also checked against config.schema.json, though during the write rather than before it: the staged assets are copied into the study first, so a schema failure leaves a partly-populated directory rather than nothing. See the API page.

6 · Compute¶

Runs the analyses, streaming the log into the page with a running clock. It creates a virtualenv inside the study so the run does not depend on how you installed the builder.

This is the step that can take real time, and what governs it is mcCount, not the size of your study. The coherence null costs the same for an 8-second clip as for a twenty-minute recording, and results are cached, so the cost scales with how many distinct nulls are needed.

The default is 0 — no null — unless you switched the network on, which asks for 100. Only a study that sets mcCount itself reaches 300, the publication setting, which is the one that runs into hours.

7 · Open it¶

Preview the dashboard locally, and the command to deploy it. For a private study this is also where the reminder sits that the data stays out of git.

What it does not do¶

It does not produce your time series. DIMS is not a preprocessing pipeline: it assumes time-aligned modalities already exist. Tracking, filtering and synchrony estimation happen before the wizard.
It does not check that your alignment is correct — only that you were given the chance to fix it in step 3.
It is not a server. It binds to localhost, has no authentication, and is meant to run on the machine holding the data.