Admin UI

A small built-in web panel to see what Gligabao stored, what it cost, and how your models and evals are doing. It is part of the server binary: no separate service to run, and it talks to the server's own HTTP API.

What it does

Four views, one panel.

Inspect your memory

The main view lists the datasets of every user in each database, with live health badges for the server and the sidecar in the top bar.

  • Datasets and totals: per dataset, its artifacts, entities and facts, and the totals of the whole system. Tick include deleted datasets to see soft-deleted ones.
  • Graph viewer: an interactive force-directed view of a dataset's knowledge graph. Facts show when they were true in the world and when Gligabao learned them, and which model inferred each one.
  • Vector DB stats: per index (entities, facts, chunks, summaries), how many vectors are stored, how many should be, how many are missing, and the dimension.
  • Delete a dataset: a two-step delete, with no accidental one-click removal.
  • Paste an image: paste one anywhere on the page, give it a user and dataset, and press Remember. The extracted metadata appears beside it.
  • Image store and web search cards: the images tracked by type and theme, how the image directory compares with the database, and the web searches each dataset holds, with their sources and the facts they yielded.

Watch every model call

Three views for what happens inside the pipeline and what it costs.

  • Remember runs: how many runs are started, succeeded or failed. Click a status to filter, and click a run to see its parameters and how each stage went (status, time, error, inferences). A run that succeeded but has a failed stage is flagged.
  • Inferences: every recorded model call with tokens and cost, broken down by operation, stage, model and thinking level. Click a row of a breakdown to filter the list, newest first.
  • Timeseries: the series stored in each database across all users: the period each covers, its length, the entities and anchors it belongs to, and its stats.

Run training from the browser

The training view is a console for the fine-tuning loop of Gemma 4, from dataset to tested model, without the modal CLI.

  • Datasets: list the datasets on the server, upload a JSONL file (instruction, output, optional input), and open one to read its lines.
  • Jobs: pick the model (4B, 12B or 31B), the dataset and the number of steps, then press Run. Select a job to follow its loss curve and live log as it streams in.
  • Trained models: each finished job becomes a model of its own, so training again never overwrites an earlier one. Open a model to see its files, run inference on it, and browse its past requests.
  • Audit and Eval buttons: an audit asks the model the questions of its own training data, to check the training took. An eval asks it questions it never saw, to check how well it generalizes. Both run in the background, are scored by an LLM judge, and stream live, answer by answer, with an option to cancel.
  • Providers: see whether the Modal backend is healthy and the app deployed, then deploy, redeploy, stop it, stop warm GPU inference containers, view the live deploy log and list the volumes. A notice warns when the sidecar is not active.

Compare eval experiments

The evals view reads the separate eval database written by ./run_evals.sh and lets you dig into every run.

  • Experiments and runs: counts and a list of experiments, their runs, and the results of each case with its checks, passed or failed.
  • Case history: one case across all runs of an experiment, plotted as status, latency and checks passed, so you can see when a failing case started to pass.
  • Trace viewer: the ATIF trajectory of a case as a flow graph, with every step, tool call and result. Click a node to read its full content.
  • Per-case graph and searches: open the knowledge graph a case produced, and the web searches it made, including the ones that found nothing.

Turn it on

The panel is opt-in. Build the server with it embedded, then start the server with the flag.

terminal
# build with the panel embedded (needs node/npm), then serve it
$ ./run_build.sh
$ ./bin/gligabao serve --admin-ui     # or admin_ui = true in gligabao.toml

# then open it in the browser
$ open http://localhost:7878/v1/gb/admin/

Make it yours

The panel can be extended when gligabao is embedded in your own program.

Admin extensions

If you embed gligabao in a Go program, you can add your own pages and buttons to the panel with a folder of plain JavaScript, with no build step. An extension that fails is skipped and the rest of the panel keeps working.