CEDX Copilot · AI
Ask AI — ⌘J — sits on every screen of every CEDX app. This is the console behind it: prompts versioned like code, every run traced, answers scored against eval gates, and spend watched against a budget.
Every answer in these captures ends the same way — “Cite handbook §Billing/double-charge.” An assistant that shows its sources is one you can check. That is the whole design.
Runs on demo data. The budget card says “illustrative” on the screen itself.
What it is
The Playground's system prompt is on screen: “Answer only from retrieved context. Cite sources.” The demo transcript does exactly that — finds order NL-48291, names invoices INV-9912 and INV-9913, applies refund policy §4.2 and ends by citing the handbook section it relied on.
The registry carries VER, project, model, runs, score and cost per run on every row — docs.rag.answer is on 4.0.0 with 22,110 runs at $0.0142 each. Tag chips filter the list by what you are actually tuning: grounded, latency, policy, cost.
320 runs with run id, prompt, model, status, latency, in and out tokens, cost and score. The failures are as legible as the successes: Error rows sit at 30.0s with zero output tokens — a timeout, not a mystery.
Product tour
Not a mockup and not a concept deck. This is what opens at /app/copilot.
01 — Overview
Docs RAG citation quality slipped after the Jul 30 traffic spike — citation F1 fell to 0.79 against the 0.82 gate. Reasoning-model spend is up 38% week over week. Each alert arrives root-caused and priced: symptom, cause, fix, and the dollar figure next to it.
02 — Playground
Template support.triage.v3 on assist-balanced, temperature 0.2, with the transcript labelled “demo transcript · no live assistant call” on the screen itself. The template score of 86 comes from 18,420 historical runs, and the screen says so.
03 — Prompts
84 prompts, 501,801 historical runs, 6 models in use, an average score of 77 across the view. Every row names its version, project, model, run count, score and cost per run — and exports to CSV.
04 — Traces
320 traces with status, latency, in and out tokens, cost and score per run. The most expensive run in view — sales.rag.hotfix at $19.48 — sits next to a 91 score, so cost and quality are argued about from the same table.
Who runs it
Roles, not references. We have no named customers yet, so nobody in these photographs is quoted, credited or claimed as one.
Owns the registry: versions prompts, watches score and cost per run, and promotes 4.0.0 over 3.0.8 only when the evals say so.
84 prompts · 6 models
Keeps the handbook the assistant is allowed to cite. Retrieval quality is their scoreboard — citation F1 against the 0.82 gate.
citation F1 · gate 0.82
Reads the traces when an answer goes wrong, and works the review queue before a template ships to production.
review queue · 23The shape of it
Every figure below is legible in the captures above. Nothing here is a projection of your estate — it is the state of the demo data.
How it runs
A prompt starts in the Playground: template, model and temperature set before the first run, with the cost estimate on screen.
Evals gate the answer. Citation F1 at 0.79 failed against the 0.82 gate, and the alert arrived root-caused rather than as a dip in a chart.
Every run lands in Traces with tokens, cost and latency, so a bad answer in production has a paper trail back to the exact run.
Versions move through the registry — docs.rag.answer is on 4.0.0 — with runs, score and cost per run following the version, not the file.
One record
Copilot is not a chatbot with a connector bolted on. Its system prompt says “answer only from retrieved context” — and that context is what the rest of the estate writes.
The retrieved context the system prompt restricts answers to is the handbook Knowledge maintains — which is why citation F1 has a gate.
CEDX DeskThe support.triage.v3 template in the captures answers exactly what Desk hears: invoice disputes, refunds, order lookups.
CEDX AI GatewaySix models in use on the registry. The gateway is where each model's card, price and quality bias live.
CEDX AgentsThe review queue and the frontier are where scored runs become supervised, longer-running work.
Limits
Finding this out on the third call is worse for you than reading it here, and worse for us.
Start
Open it right now
Your prompts, your handbook
Copilot with the rest of it
Questions
Yes. Every screenshot is a capture of the running build and you can open the same build at /app/copilot. It runs on demo data, which the Playground states on screen rather than in a disclaimer.
From retrieved context, and only from retrieved context — that restriction is the system prompt, visible in the Playground capture. The demo answer ends by citing the handbook section it used, §Billing/double-charge, so a person can check it.
The eval gates. When citation F1 fell to 0.79 against the 0.82 gate after a traffic spike, the Overview raised a root-caused alert naming the prompt, the model and the question type it concentrated in — not just a downward line.
The registry prints cost per run on every prompt — from $0.0012 on assist-mini to $4.22 on assist-large in the current view — and every trace carries its own cost. The budget card is labelled illustrative; nothing on this page is a price list.
Yes, at the run level. Traces holds each run's status, latency, in and out tokens, cost and score; the error rows sit at 30.0 seconds with zero output tokens, which is a timeout you can act on rather than a complaint you cannot.
No certification has been issued. What we can evidence about hosting, encryption, tenant isolation and production access is written up on the security page.
Live build, demo data, no card. Then ask it the question your support team got this morning.