Skip to content

WEB_UI.md

Latest commit

 

History

History
154 lines (112 loc) · 5.03 KB

WEB_UI.md

File metadata and controls

154 lines (112 loc) · 5.03 KB
Raw

Web UI tour

chachaml ships a Ring/Jetty web UI that reads from the same store your REPL writes to. It's not a separate service; it's a different view onto your chachaml.db (or shared Postgres).

Starting the UI

clojure -M:ui                          # default localhost:8080
clojure -M:ui --port 9090              # custom port
PORT=9090 clojure -M:ui                # via env var

Or from the REPL:

(require '[chachaml.ui.server :as ui])
(ui/start! {:port 8080})

The server reads from whatever store your REPL is using (*store* / chachaml.db by default). Switch to Postgres before starting start! to point the UI at a shared backend.

The pages

/ → redirects to /runs

There is no landing page. The first thing you want is the runs list.

/runs — Runs dashboard

Every run, newest first. Columns: id, experiment, name, status, created-by, started, duration, the headline metric (the most recently logged value of any metric).

Filter by experiment via the dropdown. Click a column header to sort. Each row links to /runs/:id.

Tip: if you log a metric called :accuracy or :f1, the runs list shows it as the headline. Use a stable name across runs in an experiment so the table is comparable.

/runs/:id — Run detail

The full picture of one run:

  • Header with status, experiment, name, tags, created-by, parent run, duration.
  • Params table.
  • Metrics — both the latest scalar values and a Vega-Lite chart per metric key (e.g. loss vs. step).
  • Artifacts — name, size, content type, sha256. One-click download.
  • Notes — markdown box; edit and save.
  • Tags — add/remove inline.
  • Datasets — anything logged via (ml/log-dataset! ...).
  • Environment — git sha, git branch, JVM, OS, captured at start-run!.

Tip: if a metric was logged with step, the chart x-axis is step. If only one point exists, you'll see a single dot — fine for final-only metrics.

/compare?ids=a,b,c — Compare runs

Pick 2–N runs from the runs list (checkboxes + "Compare" button), or craft the URL by hand. Side-by-side: params, scalar metrics, plus overlaid metric curves on the same chart.

Tip: great for confirming a hyperparameter change moved a metric in the right direction.

/experiments — Experiments overview

Rows of experiments with their run counts, last-updated timestamps, and the metadata you set with (ml/create-experiment! "name" {:description "..."}).

/search — Metric search

Free-form: pick a metric key, an operator, a threshold. Returns the runs whose latest value of that metric satisfies the predicate.

Equivalent of (ml/search-runs ...) from the REPL.

/models — Model registry

All registered models with their current production version, description, and the version count.

/models/:name — One model's history

Versions table with stages, descriptions, source run links, and a one-click promote button. Promoting to :production instantly demotes the previous production version (see ADR-0006).

The diff button between any two versions opens a side-by-side of the artifact metadata, params from the source runs, and metrics.

/chat — Chat with your data

Natural-language Q&A. Pick a provider (Anthropic or OpenAI), paste an API key (stored in your browser's localStorage, never sent to the chachaml server), ask a question. Behind the scenes the page POSTs to /api/chat, which calls chachaml.chat/ask.

Example questions:

  • "Which run in 'iris' has the best accuracy?"
  • "Compare runs abc and def."
  • "List the production models."

The model uses the MCP-equivalent tool set to query your store; it never sees the data unless it asks for it.

The JSON API

Every UI page has an underlying JSON endpoint, useful for scripting and debugging:

Read Endpoint
List runs GET /api/runs?experiment=...&status=...
Get run GET /api/runs/:id
Compare GET /api/compare?ids=a,b
Search GET /api/search?metric_key=accuracy&op=>&metric_value=0.9
Export CSV GET /api/export?experiment=...
List models GET /api/models
Get model GET /api/models/:name
Diff versions GET /api/diff/:name/:v1/:v2
Experiments GET /api/experiments
Artifact download GET /api/artifacts/:id/download
Tags GET /api/tags/:id · POST /api/tags/:id
Note POST /api/note/:id
Datasets GET /api/datasets/:id
Chat POST /api/chat

There's also a write API (/api/w/...) for non-Clojure clients: start-run, log-params, log-metrics, end-run, log-artifact, register-model. See log-from-python-or-curl for an end-to-end example.

Production deployment

The UI is meant for trusted networks. There is no auth.

For team use, run the UI behind a corporate VPN, an SSO proxy (oauth2-proxy, Cloudflare Access), or in a private subnet. The team-deployment-docker how-to has a docker-compose.yml that puts the UI on a private port.