How to Connect Your Company Docs to Dust and Ask Questions
Dust is a workspace for building AI assistants that answer from your company’s own documents. You connect data sources — Notion, Google Drive, Slack, Confluence, GitHub — and then create assistants with scoped access to specific folders rather than the whole company.
That scoping is the whole point. A general chat model answers from the internet. Dust answers from last quarter’s spec doc, and cites it.
What do you actually get on the free plan?
Dust’s free tier allows a small number of users and a limited monthly message allowance — enough for one person to validate the setup, not enough for a team to lean on daily. Paid workspace plans raise both limits and add SSO. Check current pricing before you plan a rollout; it has changed more than once.
The unit that matters is the workspace. Everything lives inside one: data sources, assistants, members.
How do you connect a data source?
- Open your workspace and go to Data sources.
- Pick a connector. Google Drive, Notion, Slack, and GitHub are first-class; generic web crawling and file upload exist for the rest.
- Authorise with an account that can already see the folders you care about. Dust inherits those permissions — if the authorising account can read HR documents, so can every assistant built on that source.
- Choose folders explicitly rather than syncing the entire drive. Sync time and token cost both scale with volume.
- Wait for the initial crawl. A few hundred documents take minutes; 20,000 takes hours.
Point 3 is where companies get hurt. Use a service account whose access is narrow, or spend the afternoon discovering that your assistant will happily quote salary bands.
How do you build the first assistant?
Create an assistant, then give it three things: a name, instructions, and a data source selection.
Instructions matter more than data. A workable instruction block for a support assistant:
- Answer only from the selected sources.
- If the sources do not contain the answer, say so and name the closest document.
- Always include the document title in the reply.
- Do not speculate about pricing or roadmap.
Without the second and fourth lines, the model fills gaps with plausible text. Retrieval-augmented answers fail quietly, not loudly.
Leave the model on the default for the first assistant. Switching to a larger model improves reasoning but does not fix bad retrieval, and you will not know which problem you have until you test with the default.
How do you test whether it works?
Write 20 questions with known answers before you share the assistant with anyone. Half should be answerable from your documents, half should not be.
Score each answer for three things: correct, cited, and honest about gaps. The third column catches the most problems. An assistant that says “I could not find this in the selected sources; the closest is the Q3 pricing memo” is more useful than one that invents a number confidently.
Expect the first pass to get roughly 60 to 70 percent of answerable questions right. Most failures trace back to a document that was never crawled, or one that is scanned as an image with no text layer. Fix the source, not the prompt.
Where does Dust fit and where does it not?
It fits teams that already keep knowledge in Notion or Drive and are tired of pasting context into ChatGPT. It also fits support and sales teams who need a shared, permission-aware answer bot.
It does not fit if your knowledge lives mostly in Slack threads — Slack search is shallow and old threads often never get indexed. It also does not fit one-person setups, since the free limits and the setup effort only pay off across a team.
And it is not a document-writing tool. It answers questions. If you want it to draft a proposal from your templates, understand that retrieval gives it facts, not drafting skill — the output will need real editing.
Quick Summary
- Dust builds AI assistants that answer from connected company data, with citations.
- Connect data sources with a narrowly permissioned account; assistants inherit account access.
- Sync specific folders, not the whole drive — crawl time scales with volume.
- Write the assistant instructions to refuse speculation and to name the closest document.
- Test with 20 known-answer questions before sharing; a 60 to 70 percent first pass is normal.
