Home / Documentation / AI Assistant

AI Assistant

The AI Assistant is not a standalone chat tool but an agent that operates the platform itself: it wraps devices, products, commands, time-series data, the system and users into tools. You describe a goal in natural language and it picks tools, reads data, gives conclusions and sends commands when needed.

5 sections · Updated with each platform release

Start a conversation

Jiwu AIoT Platform · [object Object] Interface mock
ProductsDevicesData ScreensAI AssistantCredentials
Did the air compressor in Workshop 3 show any pressure anomalies yesterday?
Tool call Device data query · product=Air compressor type A · time=yesterday
Two readings exceeded 0.85 MPa:
· Workshop 3-Air compressor-01　13:42　peak 0.91 MPa
· Workshop 3-Air compressor-02　17:08　peak 0.88 MPa
Both lasted under 5 minutes and then fell back on their own.

Every AI tool call can be expanded to inspect its input and output — where the conclusion comes from is transparent.

[object Object]
  1. 1[object Object]
  2. 2[object Object]
  1. Open the "AI Assistant" and describe your goal in the input box, including scope and time where possible (for example "yesterday", "Workshop 3", "offline for more than an hour").
  2. The AI shows which tools and parameters it is calling; expand them to see exactly what it queried.
  3. Check whether the conclusion matches your expectation. If not, add conditions and ask again — that works better than starting over.
  4. File useful conversations into groups so they can be reused and found later.

NoteAI conclusions come from real platform data, but the AI does not guess: when conditions are vague it may ask back, or state plainly that the data is insufficient. That is normal behaviour, not a fault.

Which capabilities the AI can call

  • Devices: query device lists, device status and live or historical device data.
  • Products: query product lists and the devices under a product.
  • Commands: send commands to a given device (subject to account permissions).
  • Conversation insights: summarize and search existing conversations.
  • System and users: query system configuration and user information to answer questions such as "who can see which data".

NoteThe AI calls tools with the permissions of the current account, so data you cannot see the AI cannot see either — using the AI never grants extra access.

Models, prompts and skills

  • Model vendors and models: configure vendors and keys in system management and choose available models; models differ in speed and quality.
  • Prompt templates: turn frequently used ways of asking into templates so the whole team asks consistently.
  • Skill box: write "how to do things" as skills, synchronized from a skill repository, uploaded as packages or connected to an MCP catalog, with revision history kept.
  • Toolbox: the catalog of built-in tools, queryable and auditable, used to confirm exactly which capabilities the AI has.

Turn a conversation into a scheduled job

  1. Create a job under scheduled jobs in the "AI Assistant".
  2. Enter the question to run periodically, for example "count all devices offline for more than an hour today, grouped by product".
  3. Set the execution interval and time.
  4. After saving, the execution history shows the input, output and duration of every run.

NoteScheduled jobs suit daily-report style inspections. Alarms that need millisecond real-time behaviour should not be built in the AI; use the platform’s alarm and command linkage capabilities instead.

How to ask for better results

  • Include a time range: yesterday, this week, the last 7 days.
  • Include an object scope: a product, a workshop, a specific device.
  • Ask for structured output: group by product, list the top 10, give a one-sentence conclusion.
  • Ask for a report: have the AI produce an inspection summary that can be forwarded directly.
  • Ask one thing at a time; splitting a compound question into two rounds noticeably improves accuracy.