Quick Start

From zero to a working agent on your own data in under an hour. This walks through the typical first day.

Before you start

You need:

A StellarBase account (managed EU cloud — free trial available) or a self-hosted deployment
At least one data source you want to connect (Google Drive folder, SharePoint library, Postgres DB, or a set of PDFs to upload)
Some idea of the first workflow you want to enable (lit review, contract check, policy review — pick something concrete)

If you haven’t picked a deployment mode yet, read Deployment Overview. For a trial, managed EU cloud gets you running fastest.

Step 1 — Create a Base

A Base is the top-level container. It has its own data, its own agents, its own permissions. For a first pilot, create one Base dedicated to the task you’re trying.

Give it a clear name (“Pilot: Contract Review”, “Research — Surface Codes”, “Oncology Tumor Board”) and set language preferences that match your content.

Step 2 — Connect a data source

Open the Base, click “Add source”, and pick a connector. The UI walks you through OAuth or credentials as needed.

Common first connectors:

Google Drive / SharePoint — a single folder is enough. Don’t try to index everything on day one.
Upload — drop a folder of PDFs or Word docs directly into the UI
Postgres / MySQL — point at a specific table or view if you have structured data

Ingestion is asynchronous. A few hundred documents typically finish in a few minutes. A few thousand can take up to an hour. Progress is shown in the Base.

See Data Sources for the full connector catalog.

Step 3 — Explore via chat

Once the first batch of documents is indexed, open Chat and ask something specific. Not “summarize the database” — that’s too broad. Try:

“Find all contracts mentioning change-of-control from 2024”
“Which of these papers discuss [specific technique]?”
“Summarize the three most-cited documents about [topic]”

Every response cites its sources. Click citations to verify. This is the fastest way to see whether the indexing worked and whether the corpus covers what you need.

Step 4 — Create your first agent

The magic moment. Agents encode your playbook. Click “Create agent” and fill in:

Name — descriptive (“MSA Reviewer”, “Literature Screener”)
Role / description — one sentence
System prompt — this is where your domain knowledge goes. Start simple: “You are a [role]. For every input, do [task]. Flag [criteria]. Always cite the source document + page.”
Knowledge scope — which collections the agent can read
Tools — start with search + OCR; add more as needed
Output schema — optional, but recommended for downstream consumption

Test it on 3–5 documents you know well. Refine the prompt based on what it gets wrong. This iterative loop usually takes 15–30 minutes per agent before it’s production-ready.

See Agents for detailed prompt-writing guidance.

Step 5 — Wire a workflow (optional for now)

For recurring tasks, turn your agent into a workflow. The visual editor lets you chain nodes:

Input — new documents in a folder, webhook, schedule
Processing — StellarOCR, entity extraction, anonymization
Agent invocation
Output — CSV row, database insert, email, Slack message

You don’t need this on day one. Day one is about proving the agent works. Day three or five is when workflows pay off.

See Workflows.

Step 6 — Invite colleagues

Add teammates to the Base with the appropriate role (editor for people configuring agents, reviewer for people validating outputs, viewer for people just consuming). Everyone gets their own chat thread, sees shared agents and workflows, and accumulates the shared audit log.

See Collaboration.

Common first-day pitfalls

Trying to index everything at once

Start with a focused corpus — one folder, one matter, one project. You’ll get better results from 200 relevant documents than from 50,000 mixed ones. Expand later.

Starting with too broad a prompt

Agents do best on narrow, well-defined tasks. “Review this contract” is too broad; “Identify change-of-control clauses and check against our firm’s playbook” is right-sized.

Skipping citations

Always ask the agent to cite. Check a few citations manually. Trust is built on verifiability.

Assuming the prompt is done on the first try

It isn’t. Expect 3–5 iterations before an agent reliably does what you want. That’s normal — and it’s much faster than training a classifier from scratch.