Work that outlives the request that asked for it. Ingesting one
corpus was measured at 13 minutes inside a single call, against a token that
lives 15 — it now runs here instead. A queued job cancels at once; a running
one at its next checkpoint.
Kind
Status
Progress
Started
Duration
Outcome
One row per RAG profile: whether its corpus is ready, when it was
last ingested, the grounding threshold below which it refuses, and what it has
cost. A profile with a threshold of zero answers whatever it retrieves.
Profile
Category
Adapter
Rows
Threshold
Last ingest
Queries
Tokens
External
Warnings
—
Three tables described a training run and none knew about the
others, so nobody could say what had been trained here. This reads all three
into one listing. A YOLO run started before this existed left no row at all —
it lived in process memory and was lost at restart.
Branch
Base model
Status
Progress
Started
Duration
Metrics
Source
Run detail
Pick a run to see its parameters and dataset snapshot.
Twenty-nine endpoints meant "populate the defaults", each with
its own name and its own answer. Initialising a tenant meant knowing all of
them and the order. These run in dependency order and never overwrite an
admin's edit: a second run creates nothing.
Target
Branch
What it populates
Needs first
Scope
Last run
Result
Run a target to see what it created and what it left alone.
Accuracy and leakage are measured separately because they fail
separately. A run with good accuracy and non-zero leakage is a failed run:
leakage counts the out-of-scope questions the assistant answered instead of
refusing. Target is zero.
History
Started
Trigger
Passed
Pass rate
Out of scope
Leaked
Leakage
Last run detail
—
Where the time goes and which source answers. An average hides
the thing you need: the audited instance reported a mean of 46.5 seconds over
64 queries, which says something is wrong and nothing about what. Percentiles
say whether it is every query or a long tail.
Latency
Which source answers
Route
Queries
Share
Providers
Provider
Queries
Share
Top intents
Intent
Queries
Recent answers
When
Question
Category
Route
Provider
Confidence
Latency
What the assistant actually did: which category it routed to and
by what margin, what evidence it retrieved and with what score, which tools it
reached, and whether it refused before calling a provider. Diagnosing a weak
answer used to mean reading the logs.
Answer
—
Retrieved evidence
Source
Category
Score
What
Full payload
—
Spatial tools available to the assistant
What the geospatial side can do when a question reaches for it.
These take GeoJSON in and give GeoJSON out — the assistant still needs
something to hand them, and today no answer path produces geometry to feed
them. Listed so the capability is visible rather than buried.