Ask rsync.ai
A chat panel inside the product. Ask a question in plain words and it answers from your connections, schemas, pipelines and runs, or drafts the thing you want. It reads metadata by default, not your rows.
Open it
Select Ask rsync.ai in the header, or press ⌘/ (Ctrl+/ on Windows and Linux). On a pipeline, connection, run or dashboard page the question is about that page; the panel shows an About chip you can remove. Each answer says which tools it looked at and which model answered. Stop ends a question on the server as well as in the browser.
What it can do
| Ask | What happens |
|---|---|
| “Why did this pipeline's last run fail?” | Reads the run and answers from its state and scrubbed error text. |
| “Sync customers and orders from Postgres to BigQuery” | Proposes a pipeline. You create it from the link it gives. |
| “Orders by status last month on the shop database” | Writes SQL, marked Not run, and opens it in the Explorer. |
| “Add a connector for Acme from https://api.acme.com/openapi.json” | Starts connector generation (owner or admin, Cloud). |
| “Every morning at 8, post failed payments to #data-alerts” | Drafts a workflow for you to test and turn on. |
| “A sales dashboard from the daily orders query, with a date filter” | Drafts a dashboard you preview and save. |
| “Make the orders chart a line, full width” | Proposes a one-tile change you can apply. |
It uses the same tools as the agent gateway, run as you: it sees what you can see in your active workspace, and nothing more. What it can do follows your role:
| Role | It can |
|---|---|
| Any member | Answer questions about your workspace and propose a pipeline. |
| Owner or admin | Also create, run and schedule models, generate a connector, and draft a workflow or dashboard. |
What the model sees
The model is a third-party service, so by default it gets metadata only:
- your question and the earlier turns of the conversation;
- names, types and settings the tools return: connections, schemas, pipelines, run states, and error text with literal values scrubbed out;
- your plan and how much of it is used, as counts only;
- rsync.ai's own guides on how pipelines, workflows, dashboards, models and connectors work.
It never gets query results, row counts, credentials or personal data.
Sample rows (off by default)
When names and types don't say what a column holds, the assistant can read a few rows. Two switches must both be on: a platform admin allows it for the model (Admin → Settings → AI models), and the workspace owner turns it on (Workspace settings → Data). Then it reads at most 3 times per question, 5 rows each. Columns that look personal or secret are left out, and every value is checked for personal data before it's sent. Each read is written to the audit log.
Limits
- One question gets at most 6 model turns, 12 tool calls and 120 seconds. A question that reaches a limit still gets an answer, marked as stopped.
- 20 messages a minute per person, shared with chat.
- It counts against your own monthly LLM spend limit, which an admin sets.
- Chats are kept in the browser tab (up to 20) until a reload or sign-out clears them, and the server keeps the last five turns for 30 minutes. Nothing else about a chat is stored.
Self-hosting
A self-hosted install ships with Ask rsync.ai switched off. Turn it on with one environment variable on the api-gateway service, then give it a model:
RSYNC_ASSISTANT_ENABLED=true
- It supports Claude, ChatGPT and Gemini. Each is offered once it has a key.
- A platform admin adds keys under Admin → Settings → AI models, or you set them on the
llm-serviceservice. A saved key is encrypted and never shown again. - Each person picks which of the available models to use in the panel.
Connector generation is part of rsync.ai Cloud. See Self-hosting for the full install guide.
Ready to try it?
Start building on rsync.ai Cloud today — or self-host on your own infrastructure.