Deep dive
Own one job before becoming a workspace
A blank, infinitely flexible workspace is hard to sell to a new customer. What works is a product that does one job very well and grows from there: a support knowledge base with article analytics, an agency's client portal with project docs, a research notebook with citations, a compliance wiki with review cycles and approvals.
That job decides your first block types, templates and integrations, and it gives users a reason to switch. The flexible editor and databases still matter, because they let customers adapt the tool, but the MVP should arrive with opinions. Our MVP development process keeps those opinions explicit.
How blocks and real-time collaboration work
Every piece of content is a block with an ID, a type, properties and children. A page is a block whose children are other blocks, which is why anything can be nested, moved or turned into another type. Store blocks relationally for permissions and queries, and keep a collaborative document state for each page for editing. Keeping both in step is one of the subtle parts of the design, so decide early which one is the source of truth for each kind of change.
For editing, conflict-free replicated data types, or CRDTs, let each client apply changes locally and merge them with everyone else's in any order, always reaching the same result. Yjs is the most widely used library, Automerge offers a strong change history and Loro is a fast newer option. A sync service relays updates over WebSockets and stores snapshots so pages load quickly.
- Apply edits locally first; never make typing wait for the network.
- Persist updates before acknowledging them, and snapshot long documents periodically.
- Test with several clients editing the same paragraph on slow, flaky connections.
Databases, permissions and search
Database rows are pages with typed properties, so a task list, a CRM and a content calendar can all be built from the same parts. Filters, sorts and views run on the server for large databases and on the client for small ones. Relations and rollups, which connect databases, are powerful and expensive, which is why they sit in the growth tier.
Permissions are inherited down the page tree with explicit overrides, guest access and public links. Compute effective permissions once and reuse them for page loads, search results and AI retrieval. Search indexes must update when pages change and when permissions change, or users will find content they should not see.
AI that cites its sources
AI fits naturally into a workspace. In the editor, it drafts, rewrites, summarises and translates the current page. Across the workspace, retrieval-augmented generation answers questions from pages the user can see, with citations to the exact blocks. Later, agents can update databases and pages, with approvals for changes that matter.
Trust depends on accuracy and control. Cite sources, let admins choose which spaces AI can use, keep customer data out of model training through provider settings and log AI actions. Our guide to building a RAG chatbot over company documents covers the retrieval side in detail.
Security, enterprise features and running costs
Workspaces hold plans, contracts and internal knowledge, so business customers ask for SSO, audit logs, exports, retention and data residency, and often a SOC 2 report. Design tenant isolation and audit events into the first release, then add enterprise controls as larger customers arrive.
Running costs are moderate: database and sync servers, file storage, search, AI inference and email. AI is the line that grows fastest, so meter it. Maintenance is roughly 15-20% of the build cost per year. Our application maintenance and support team typically handles the steady stream of editor edge cases that real users discover.