Dashboard

Workspace

Operationalise semantic standards with observability inspired by modern data catalogs. The Dashboard is the landing page of the reviewer console and the fastest way to sanity-check that the matcher is behaving before you commit to a batch of review work.

What the dashboard shows

Three metric tiles read live from the app state context:

Tile	Source	What it tells you
Canonical values	GET /api/reference/canonical	Size of your approved golden-record set
Dimensions	GET /api/reference/dimensions	How many semantic domains you have defined
Matcher backend	GET /api/config	Whether matching runs on embeddings, lexical fallback, or LLM

If the Matcher backend tile disagrees with what you set in Settings, your most recent PUT /api/config did not persist. Check that the API accepted the write and that the browser can reach it.

Semantic match playground

The playground is the fastest feedback loop in the product. Enter a raw value you expect to see in a source system, optionally pick a dimension, and see the ranked candidates immediately.

curl -X POST http://localhost:8000/api/reference/propose \
  -H 'Content-Type: application/json' \
  -d '{"raw_text": "Never married", "dimension": "marital_status"}'

The response carries a matches array of MatchCandidate objects, each with a score between 0.0 and 1.0, ranked best-first:

{
  "raw_text": "Never married",
  "dimension": "marital_status",
  "matches": [
    {
      "canonical_id": 3,
      "canonical_label": "Unmarried",
      "dimension": "marital_status",
      "score": 0.94,
      "description": "Never married, single, no legal partner"
    }
  ]
}

The UI renders this as a table of label, dimension, description, and confidence percentage.

The semantic matching playground: type a raw value, see ranked candidates with confidence scores.

Read the score, not just the rank

A high score means the embedding put the candidate close to your raw text. It does not mean the candidate is correct. The score is a ranking signal for humans; approval is always a deliberate act.

Interpreting a weak match

Symptom	Likely cause	Fix
Every score is low and flat	Embedding model has not downloaded, or the backend fell back to lexical matching	Check the Matcher backend tile and the API log for model download errors
A candidate from the wrong dimension ranks first	You left dimension unset, so the matcher searched the whole library	Pin the dimension in the playground
Good candidates exist but scores are below the threshold	match_threshold is set too high for your data	Lower it in Settings, or approve from Suggestions

Next steps

• Canonical Library

  ---

  Curate the golden records the playground is matching against.

  Canonical Library

• Suggestions

  ---

  Work through the queue the dashboard's scores feed into.

  Review suggestions

• Settings

  ---

  Tune thresholds, the embedding model, and LLM escalation.

  Tune the matcher