Known limitationsEarly Access

Supported files and limits

• Up to 25 MB per file.
• Supported uploads include PDFs, Word (.docx), PowerPoint (.pptx), Excel (.xlsx/.xlsm), Markdown/text, CSV/TSV, common code files, and engineering source/deck files.
• Plain text is indexed as text. It is not sent through OCR. Scanned PDFs and image files require the deployment OCR provider. When enabled, OCR stores searchable text plus source-derived fields and tables for review; unclear scans still need source verification. If OCR is not configured, these files fail with a visible status and are not searchable.
• Audio and video files require the deployment transcription provider. If transcription is not configured or no speech text is found, the file remains out of retrieval until you upload a transcript or reprocess it after configuration is fixed. Transcripts are source-derived and should be reviewed before formal use.

AI answer-quality caveats

• Answers come from your uploaded sources only - not web search or general knowledge.
• AI output can still be wrong or incomplete; always verify it against the cited sources.
• Quality depends on what your documents contain.
• Every answer cites its sources, and when the sources don't support an answer, the system declines rather than guessing.
• Early Access uses a single embedding model and dimension; switching is a later capability.

AI provider modes and disclosure

• You choose how AI runs at the organization level: platform-managed, bring-your-own-key, or local/self-hosted.
• On the platform-managed path, your source text, questions, and uploaded content needed for provider-backed OCR/transcription may be sent to the external OpenAI API to produce embeddings, answers, OCR output, or transcripts. This is opt-in; local/self-hosted is available when the review path requires stricter data control.
• Bring-your-own-key uses your own provider key (they process and bill you). Local/self-hosted runs on your own hardware, so your data stays in your environment.

AI credits and caps

• Each plan includes monthly AI credits at the organization level. Platform-managed answers, embeddings, OCR, and audio/video transcription consume this usage when those provider-backed paths are enabled.
• Managed usage is hard-capped: once you hit the cap, AI-backed requests and provider-backed ingestion can be blocked until the next period. There is no unlimited platform-paid usage.
• Bring-your-own-key and local usage are metered for visibility where supported but never platform-capped by Researcher Center. Any cost figures shown are estimates, not a bill.

Other Early Access limitations and data handling

• Sign-in is email and password by default. Google and Microsoft sign-in appear on the sign-in page only where a deployment has them configured. Enterprise single sign-on (SAML/OIDC) is in private validation and is not generally available during Early Access.
• Single region during Early Access; no multi-region failover.
• Data isolation: each group's data is isolated per tenant. We do not log your document text, prompts, answers, or secrets.
• Retention: uploaded documents and derived data are kept for the life of the workspace until you delete them; deleting removes them from the database and object store. A formal retention window is still being finalized.
• No formal uptime SLA during Early Access (best-effort; see support).

For sensitive or regulated data, use bring-your-own-key or local/self-hosted mode and confirm compliance with your institution's data-handling policies before uploading. See the privacy page.