Department scenarios

Ten agents a team can open on day one — what each does, how it is wired, and how to build your own version in an afternoon.

A department buys one box. On it sit a shared drive, a chat tool and an agent runtime. The drive already holds the work; the question is what to point an agent at first.

These ten flows are that answer — one per team, plus two routines any team runs. Each is small enough to read in a minute and open enough to change.

Every question and answer on this page is a real run against a box with an on-box model. Nothing here is an illustration.

What you can do with them

#	Team	The job it takes off someone's desk
1	Legal	Reviews a draft clause against your review guide and names the article that forbids it
2	HR	Answers a leave question by calling a calculator, so the number is computed and not guessed
3	General Affairs	Turns one line of instruction into a notice in the house format
4	Corporate Strategy	Turns meeting notes into decisions with an owner and a date
5	Finance	Picks the duplicate-payment suspects out of a payment list
6	Sales	Breaks an RFP paragraph into a requirement checklist
7	Customer Support	Grades tickets by priority and states the rule it used
8	Engineering	Turns an incident log into a postmortem stub
9	Any team	Reduces last night's mail to what needs answering today
10	Any team	Tags a batch of files by owning team and retention

Setting them up

Three steps, and only one of them is configuration.

Open Studio on the box. It is already there — the appliance ships it. Sign in with your Studio account.

Point one field at your own model. Every scenario uses the same model component, and it needs two values:

Field	What to put
Ollama API URL	the model server as seen from the Studio container, e.g. `http://host.docker.internal:11434` for Ollama on the same machine
Model Name	whatever that box serves, e.g. `qwen2.5:7b`

That field is the entire difference between an on-prem agent and a cloud one. There is no account to create, no key to paste, and nothing to count per seat.

Turn on Tool Model Enabled if the flow hands the agent a tool — scenario 2 does. Without it the model will answer from memory instead of reaching for the calculator.

Open a scenario and press Playground. Type the question, read the answer. Nothing else is required to start.

Building all ten at once

Doing that by hand ten times is a waste of an afternoon, so it is scripted:

cd deploy/platform/scenarios/studio

# create all ten through Studio's own API
python3 build_flows.py --key sk-… --model qwen2.5:7b \
    --ollama http://localhost:11434

# run all ten and print what each returned
python3 run_flows.py --key sk-…

run_flows.py prints the answer for every flow and flags any that came back with Chinese characters in Korean prose, so you can see the state of all ten in one screenful rather than clicking through them.

Change a department's instructions in build_flows.py, re-run, and you have your own version of this.

How one is wired

Legal — reviewing a clause against your own guide

Four components, and you can read the whole thing:

Chat Input ─────────────┐
                        ├─▶ On-box model ─▶ Chat Output
Prompt (review guide) ──┘

The question comes in on the left. The department's review guide sits underneath it, in a Prompt Template — not general legal knowledge, but your team's own standard, pasted in. Change the standard and the agent changes with it.

Asked

다음 조항을 검토해줘: '본 계약의 손해배상 책임에는 상한을 두지 아니한다.' 그리고 '경업금지 기간은 5년으로 한다.'

Answered

(1) 위험 여부: 위험합니다.
(2) 근거 조항 번호: 제6조, 제3조
(3) 해야 할 조치: 무제한 손해배상 조항 삭제 및 경업금지 기간 3년 이내로 수정.

Both clauses caught, each against the article that forbids it, with the action to take.

HR — a number the model is not allowed to guess

This one is worth reading even if your team is not HR.

The company rule is 15 days after one full year, plus one day for every two years of continuous service beyond the first, capped at 25. Applied to three years of service that gives 16.

Asked to work it out, the model answers 20. At temperature zero it answers 20 every time — confidently, and wrong.

So the flow does not let it. Six components instead of four:

Chat Input ──────────────┐
Prompt (rule + refusal) ─┤
On-box model ────────────┼─▶ Tool-calling agent ─▶ Chat Output
CalculatorTool ──────────┘

Two of those feed the agent for different reasons: the model is its brain, the calculator is its tool. The system prompt states the rule and forbids the arithmetic, so the agent has to reach for the calculator to answer at all.

Asked — 3년 근속한 직원의 연차휴가는 며칠인가요?

Answered

3년 근속한 직원의 연차휴가는 16일입니다.

The Playground shows the CALCULATOR calls above the answer. The agent read three years out of the question, built the expression, called the tool, and reported what came back. None of that number was invented.

Retrieval grounds finding. It does not ground reasoning. A model reading the right policy will still derive the wrong number from it. Anything that must be computed — entitlements, prorations, deadlines, totals — belongs in a tool. That holds for every department, not just HR.

Two more, briefly

Corporate Strategy — meeting notes to decisions

Asked — meeting notes where two items got an owner and a date, and one did not.

결정사항	담당자	기한
파일럿 후보 3곳 선정	김대리	9월 5일까지
제안서 초안 작성	이과장	9월 10일까지
예산 재검토 논의	미지정	미지정

The third row is the point. The notes never assigned it, so the flow says so. A summariser that fills that cell to look complete is worse than no summariser.

Finance — duplicate payment check

Asked — five payments, one rule: same supplier, same amount, twice inside thirty days.

대한물산 / 120만원 / 8/03, 8/19 / 중복 지급 의심 건

One suspect, and only one. 서울테크 also repeats at the same amount — but fifty-two days apart, so it is left alone. The rule was applied, not approximated.

How the box reaches these flows

Everything above happens inside Studio, which leaves a fair question: what has any of it to do with the box a department bought?

A member never opens Studio. They open the box's own console, go to AI Workspace, and press Run on a routine. That routine reaches a flow you built:

Member ─▶ Box console ─▶ Agent adapter ─▶ Studio flow ─▶ On-box model
          AI Workspace   nufi-app :8902   the one you built

The adapter maps a routine to a flow. It lives in deploy/platform/adapters/meshbox-agent/ and is configured with one variable:

NUFI_AGENT_URL=http://nufi-agent:7860
NUFI_AGENT_API_KEY=sk-…
NUFI_AGENT_FLOW_MAP='{"r1":"<legal-flow-id>","r2":"<strategy-flow-id>"}'

A routine with no entry in that map is honestly unwired — the adapter answers 502 rather than inventing a result.

Proving it, rather than asserting it

Press Run on 부서 문서 QA봇 in the box console, then read the run back:

1. 위험 여부: 없음
2. 근거 조항 번호: 가이드에 없음
3. 해야 할 조치: 가이드에 없음

Three parts — 위험 여부, 근거 조항, 조치 — which is the Legal flow's own output shape, and nothing else on the box produces it. The shape of the answer is the proof of which flow ran.

Two gaps, named rather than hidden.

The console reports the routine and its status and never shows what came back — a member pressing Run learns only that something happened. The result has to be read from POST /api/v1/agent/routines/{id}/run.

And the box sends the routine's name and nothing else, so a flow that needs material asks for it. Handing a routine its meeting or its folder is the next piece of work, and it belongs to the box rather than the model.

What these flows deliberately do not do

They do not mask personal data. PII masking is a real feature of the platform, but it belongs to the guardrail layer, not to a prompt. Asked to mask a Korean document, a prompt left an email address untouched and dropped an account number entirely. Deterministic detectors handle this; a model asked nicely does not.

They do not answer past their material. Each system prompt names what the flow may use and tells it to say so when the answer is not there. That is why the Strategy flow writes 미지정 and the Sales flow writes 조건부 rather than guessing.

They are pinned, not sampled. Temperature is zero and top_k is one. Left to sample, the same Korean question came back finished in Chinese — the notice-drafting flow produced 정보资产安全及保密性을. Greedy decoding removes the reach for those tokens, and makes a run repeatable enough to quote in a document like this one.

Adapting one to your team

Open build_flows.py and look at any entry in SCENARIOS. Each is six fields:

Field	What it is
`id`	the short key the adapter's flow map refers to
`name`	the title shown in Studio
`desc`	one line for the flow list
`system`	your team's standard — the part worth spending time on
`ask`	the question used to check it still works
`kind`	`prompt`, or `tool` to attach a calculator

The system field is where a scenario becomes yours. Everything else is scaffolding. Write the rule the way your team would explain it to a new hire, say plainly what to do when the answer is not in the material, and re-run build_flows.py.

Connect a model — the same model field, pointed at NuFi's gateway instead of a box.
Publishing a flow — the run endpoint the adapter calls, and who can hold a key for it.
Knowledge — when a scenario should answer from the shared drive rather than from its system prompt.

Department scenarios

On this page