Testing and monitoring

A prompt and tools that work today can break tomorrow: a page changes, an API returns a new shape, someone rewords the prompt. These are the checks that catch it, before visitors do.

Before you change anything: a set of real questions

Keep a file of questions visitors really ask, with what a good answer must contain, and run it after every change:

[
  { "question": "What's your phone number?", "source": "/contact", "contains": ["9876 5432"] },
  { "question": "Where is order 1042? I ordered with [email protected]", "contains": ["Thursday"] },
  { "question": "Do you have waterproof boots in a 10?", "contains": ["$"] },
  { "question": "Can you fix my car?", "refuse": true }
]

helppuff eval golden.json --min 0.9

Each question is asked in a fresh conversation, as a new visitor would. A question passes when the knowledge base found the expected page (source, a part of its address), the reply contains every contains, and a refuse question gets "not sure" instead of a guess. The command fails (exit code 1) below --min, so it can run in CI or before every helppuff deploy.

For tool flows, put everything in one question ("order 1042, with [email protected]"), so the assistant can call the tools straight away. Use test values your API answers the same every time (like Shippo's SHIPPO_TRANSIT).

Where to get the questions: the dashboard's Analytics → Latest questions, and the unanswered questions on each conversation's summary.

Testing one tool

helppuff tools test order_lookup --arg order_number=1042 --arg [email protected]
helppuff tools test nearest_store --prechat postcode=3056
helppuff tools test track_shipment --data order_lookup.carrier=shippo --data order_lookup.tracking_number=SHIPPO_TRANSIT

It calls the API now and shows exactly what the assistant would get (after the picked fields). On the dashboard, Test in the tool's window does the same.

Testing a whole conversation

helppuff ask "Where is order 1042? I ordered with [email protected]"    # one answer, with its sources
helppuff chat                                                        # a conversation, in the terminal

And once on the real site, as a visitor, on a phone: the widget, the forms and the tools together.

Watching it in production

Where	What it tells you
Prompt & tools, each tool (or helppuff tools list)	Its last status (200, 500, timeout) and when
A conversation's Data from tools	What every tool returned in that chat; a failure is kept as { "error": "HTTP 503" }
A conversation's summary	unanswered questions, outcome, leadQuality
Analytics	Conversations, leads, top pages, the questions visitors open with
helppuff doctor	Every piece of the deployment, with fixes
GET /api/v1/tools	Each tool's lastStatus and lastAt, for your own monitoring to poll

A habit that pays: once a week, read five conversations end to end, and add any question that went wrong to golden.json.

Changing things safely

• The prompt has versions. Prompt & tools → the prompt → History compares any version with the live one, and Restore brings one back.
• Export before a big change: helppuff agent export before.json (or Download agent.json). Importing it puts the prompt and tools back.
• Test a new tool before naming it in the prompt. A tool the prompt does not name is never offered to the assistant.
• Use a second site to experiment (a sites entry of its own), then import the agent file into the real one.

See also: [Tools]], [Agent files.