What Rabi can do for you
Rabi is the console's AI assistant. Say what you want done and it looks things up, changes settings, tests and diagnoses — without you moving between pages.
Rabi is the AI assistant in the bottom-right corner of the console. Describe what you want in your own words and Rabi looks up data, changes settings, tests agents and tracks down problems. It can also take you to the right page, or ring the button you need to press.
Refer to agents, knowledge bases and keys by name — "Support Assistant", "Product Manual". Rabi finds the item itself; you never need an ID.
Approval modes
For anything that changes data, Rabi prepares the change and runs it according to the approval mode you choose. Switch modes below the Rabi input box.
| Mode | Behavior |
|---|---|
| Confirm each (default) | Every change first shows an approval card listing what will change, and runs only after you approve. |
| Auto | Free changes run directly; anything that costs money (such as renting a GPU) still asks. |
| Plan | Look things up and plan only; nothing is changed. Useful for understanding the current state first. |
Rejecting an approval card changes nothing. Rabi reports a change as done only after it has actually run.
Examples
Build a support agent from scratch
You say: "Create a support agent called Support Assistant that answers from the Product Manual knowledge base, in a friendly tone. Then ask it how returns work to test it."
Rabi:
- Looks up your knowledge bases and finds Product Manual.
- Shows an approval card: create Support Assistant, with the instructions, base model and knowledge base listed.
- After you approve, creates the agent and asks it "How do returns work?"
- Reports the agent's actual answer and which passages of the Product Manual it cited.
You see: one approval card, a confirmation that the agent was created, and the test answer. If the answer is weak, Rabi goes on to explain the likely cause.
Find out why an agent answers poorly
You say: "Why can't Support Assistant tell me the warranty period?"
Rabi:
- Reads Support Assistant's settings: connected knowledge bases, passages retrieved and relevance threshold.
- Searches the Product Manual for "warranty" to check whether the content exists and can be found.
- Looks at recent requests about warranty to see which passages were actually retrieved.
- Explains the cause and proposes a fix — for example, "The warranty terms are in After-sales.pdf, which failed to process. Shall I reprocess it?"
You see: a specific cause and a suggested fix. When you agree, Rabi prepares the fix and shows an approval card.
Limit an API key
You say: "The Production key: at most 60 requests a minute, and only Qwen3.6."
Rabi:
- Finds the key named Production and reads its current limits.
- Shows an approval card listing the limits to apply: 60 requests per minute, Qwen3.6 only.
- After you approve, applies them and reports the result.
You see: the approval card and a confirmation once applied. Rejecting the card changes nothing.
Ask where a button is
You say: "Where do I upload documents?"
Rabi: rings the Upload documents button on the current page, with a one-line note. If the button is on another page, Rabi takes you to the knowledge base page first and then rings it.
Check usage and spend
You say: "How much have I spent this month, and which agent used the most?"
Rabi: totals requests, tokens and spend for the current workspace since the 1st of the month, ranked by agent and by model. The numbers match the Usage page.
Fix failed documents
You say: "Did any Product Manual documents fail? If so, retry them."
Rabi:
- Lists the failed documents and why they failed.
- Shows an approval card: reprocess these documents.
- After you approve, starts the reprocessing and tells you where to watch progress.
Account, usage and spend
- Account overview: plan, wallet balance and current workspace.
- Usage and spend over a period, broken down by model or agent.
- Machine rental costs.
- Open the top-up window.
"How much did I spend this month, and on which model?" "What's my balance? Open the top-up window"
Models
- Browse the model catalog: price, context length, and support for tool calling or image input.
- Add a model to My models.
"Which cheaper models support tool calling?" "Add Qwen3.6 to my models"
Building and configuring agents
- Create an agent: name, instructions, base model, and the knowledge bases and databases it uses.
- Change an agent's settings:
- name and instructions;
- base model and engine;
- randomness and maximum answer length;
- how many knowledge-base passages to retrieve, and the relevance threshold;
- memory, guardrails, and enabled or disabled.
- Connect or remove knowledge bases, database connections, API tools and MCP servers.
"Create a support agent that answers from the Product Manual" "Limit Support Assistant's answers to 500 words" "Connect the orders database to the Sales Analysis agent"
Testing and diagnosing agents
- Ask an agent a question and see its actual answer and the sources it used.
- Diagnose a poor answer: data not found, data not connected, or a model or settings issue.
- Recent requests, the full trace of one request, and failed requests with their causes.
- Content blocked by guardrails.
- Conversations held through the Agents API.
"Why can't Support Assistant answer the returns policy?" "Which Sales Analysis requests failed yesterday, and why?" "Ask Support Assistant: how is shipping calculated?"
Knowledge bases
- Create a knowledge base.
- Document list and processing status.
- Search a knowledge base to check whether a passage can be found.
- Reprocess failed documents, or delete documents.
"Which Product Manual documents failed? Retry them" "Search the Product Manual for 'warranty'"
Database connections, API tools and MCP servers
- List database connections, API tools and MCP servers.
- Test a database connection.
- Inspect the tools an MCP server provides and its current status.
- View an API tool's configuration.
"Can the orders database be reached?" "Which tools does this MCP server offer?"
API keys
- List API keys and their limits.
- Create an API key.
- Set rate limits: requests per second or minute, and token caps per hour, day or longer.
- Restrict a key to certain models, or lift the restriction.
For security, Rabi never shows a full key. Disable, delete or regenerate keys on the API keys page; Rabi can take you there.
"Add a 60-requests-per-minute limit to the Production key" "Create a key that can only use Qwen3.6"
Inference endpoints, deployments and fine-tuning
- List inference endpoints, and start or stop them.
- Deployed apps and their logs.
- Fine-tuning job progress and results, or cancel a running job.
"Which inference endpoints are running?" "Why didn't the latest deployment start? Show me the logs"
GPUs and machines
- List your machines.
- Search rentable GPUs by price or spec, and submit a rental request.
- Rental request status.
"Find machines with a GPU under $2 an hour" "Has my rental request from last week been approved?"
Operating the console
- Go to any page, or back.
- Ring the button you need: ask "where is…" or "which button…" and Rabi marks it on screen.
- Switch workspace, interface language, and dark or light theme.
- Open notifications, the command palette, linked accounts and the top-up window.
- Restart the guided tours.
"Where is the upload button?" "Switch to my organization workspace" "Change the interface to Japanese"
Documentation
- Search the documentation, with the source page linked.
"How do Agents API sessions work?" "How do I make an agent answer only from its knowledge base?"
Tips
- State the whole goal at once. "Connect Support Assistant to the Product Manual, then ask it about returns to test" — Rabi does the steps in order.
- Names are enough. A partial name works; when several match, Rabi lists them for you to choose.
- Ask from the relevant page. Rabi uses the page you are on, so "this agent" or "this knowledge base" means the one in front of you.
- Use your own language. Rabi answers in the language you ask in: Traditional Chinese, English or Japanese.
Limits
- Rabi performs only the actions on this page. For anything else it says so and takes you to the page where you can do it.
- Rabi has exactly your permissions: what you cannot see or change in the console, Rabi cannot either.
- Concurrent Rabi and agent conversations are capped; at the cap a new conversation waits briefly and shows that the service is busy.
Quick Start: Build an Assistant That Uses Your Data
One walkthrough — pick a model, build a knowledge base, connect a database, add an API tool and an MCP server, bind them to a custom model, and verify in Playground.
Quick Start: Connect the API to Your System
Create a key, choose which models it may call, set rate limits and quotas, send the first request, then trace every call in the logs.