Projects and files
Organize work, upload evidence, understand extraction states, and keep each decision isolated.
One project, one decision boundary
A project is the ownership and memory boundary for its documents, runs, artifacts, and conversation history. Keep unrelated facilities or decisions in separate projects so evidence and assumptions cannot bleed together.
Anonymous trial projects are scoped to an unguessable browser cookie. When the visitor creates an account, those projects are claimed into the verified user account and remain private to that owner.
Upload and extraction behavior
- Each accepted file is capped at 25 MB and stored durably with the project in production.
- PDF, CSV/TSV, XLSX, DOCX, PPTX, JSON, and text-oriented evidence have dedicated or fallback extraction paths; the returned status is more important than the extension.
- Usable means evidence was extracted. Pending means the file is saved but extraction has not produced evidence yet. Failed extraction means the file exists, but its contents cannot safely support analysis.
- Scans, charts without machine-readable values, protected files, and unusual office formats may require the governing pages or tables as text or CSV.
Upload success is not evidence success
Do not treat a saved filename as proof that the agent read its contents. Use the upload status, extracted highlights, and stated limitations.
History and continuity
Signed-in projects, uploaded bytes, agent runs, run events, and conversation snapshots are stored in the account-scoped workspace database. That allows the same work to survive application restarts and continue on another signed-in device.
The run record—not a browser-only transcript—is the source of truth for completed agent work. Draft and legacy chat snapshots support UI continuity but cannot silently replace a run result.