How your documents are processed
The private workspace stores uploaded files and prepares their text for retrieval. Finding relevant passages and generating an answer are separate steps. Depending on the route you use, document text or retrieved passages are processed by AI providers. This page explains those product workflows and the boundaries to check before adding internal material.
From a file to a source-backed answer
- Upload. Original files are stored in Supabase Storage, associated with the workspace. Uploading to the private workspace is a server-side operation, unlike the browser-local marketing demo.
- Extract and index. The application extracts readable text and stores smaller passages for retrieval. Text is sent to OpenAI to create the search representations used to find relevant passages.
- Retrieve. A question or search is used to locate material from the available document collection. A retrieved passage is a portion of the source, not a complete reproduction of every file.
- Answer. Built-in Ask sends the question and selected source excerpts to Anthropic. A connected external assistant receives retrieved material through its connection and uses its own answer-generation workflow.
These steps mean that workspace processing is not entirely on your device. Do not treat a private library as a promise that its contents never reach a processing provider. Check your organization’s rules for each source and the assistant account you intend to use.
Which information is used by which service?
| Service or route | Information involved | Purpose |
|---|---|---|
| Ask Your Docs and Supabase | Uploaded originals, extracted text, retrieval passages and document records. | Store and organize the library, process files and return relevant sources. |
| OpenAI embeddings | Text submitted for the search-indexing and retrieval workflow. | Create numerical search representations for matching relevant text. |
| Built-in Ask through Anthropic | The question and selected named source excerpts. | Generate an answer grounded in the supplied passages. |
| A connected external assistant | Material returned through the enabled connection, along with the conversation you have in that assistant. | Use the retrieved sources in the assistant’s own workflow. |
The way an external provider handles a conversation also depends on its service, your account and its configured settings. An Ask Your Docs connection does not replace those settings or the provider’s terms. Avoid assuming that policies for one assistant account apply to another.
Built-in Ask and external connections are different
Built-in Ask requires the signed-in user’s Anthropic API key. The saved key is encrypted on the server and used for that user’s answer requests. It is not a shared team credential. This key requirement is separate from having access to an external assistant such as ChatGPT or Claude.
An external connection lets an assistant retrieve available source material. Follow the setup for that assistant and verify an answer with sources. A setup link, copied command or authorization screen alone is not proof that the assistant has retrieved the intended files.
Before inviting someone into a workflow, review the relevant workspace and sharing controls. A source label, filename or Markdown metadata field does not enforce permission by itself. Use the actual controls for the document collection and the assistant account.
What file extraction does and does not preserve
- PDF: readable text is extracted. Image-only scans need a usable text layer; the workspace extraction path does not perform OCR.
- Word: DOCX text is extracted. The library is not a Word editor and does not preserve the complete visual layout or review history as an editable document.
- Excel: XLSX text and saved cell values are extracted. Formulas are not recalculated, and extracted passages do not retain spreadsheet cell coordinates or a working calculation model.
- PowerPoint: PPTX slide text is extracted. Do not rely on this to interpret diagrams, speaker notes or a slide’s complete visual meaning.
- Markdown, text and CSV: readable file text is used. A link inside a Markdown file does not automatically import its target.
Keep the original available when layout, numbers, visual information or an approval record matters. A returned passage may support one statement while omitting another important part of the document. Compare the answer with the source before acting on it.
Updating and deleting documents
When you replace a document, verify that subsequent questions use the intended version. The source’s name alone is not enough: check a fact that changed and the passage supporting it. Older stored file versions may remain after replacement; a successful update does not prove immediate removal of all earlier copies.
Deleting a document removes its workspace document record and requests removal of associated current files and ingestion records. Do not interpret that action as immediate erasure from every backup, older stored version or external assistant conversation. This page does not state a guaranteed retention or deletion timeframe for those separate copies.
If a document has already been retrieved into an external assistant, deleting the library source does not automatically delete the assistant’s conversation history. Review that assistant’s own controls and any copies your team exported or shared.
The public demo and free converters
The homepage demo uses a supplied example. Selecting an optional file there previews its name, type and size; it does not add that file to a private workspace or connect it to an assistant. Sign-in is the route into the actual library.
PowerPoint AI summaries
The PowerPoint summarizer reads the original file locally. Choosing Generate sends the selected slide text and numbers through our server to OpenAI. Speaker notes are sent only if selected. The request excludes the original filename and file bytes. Our summary service does not persist the source text or answer; OpenAI processes that text under its applicable API terms.
A signed, anonymous browser cookie supports the daily free allowance. The server stores keyed daily identifiers and counters in Supabase, not raw IP addresses or document content. Old counters are pruned on subsequent requests after their usage date is more than two days old. Cancelling in the browser cannot guarantee cancellation of an already-started provider request.
Local converters
The Excel, Word and PDF converters process selected files in the browser and provide local previews and downloads. They do not upload those files into the private library. A downloaded result, clipboard copy or later workspace upload is a separate action.
The public site uses analytics for visits and interactions. The custom marketing click events do not include selected file contents, filenames, email addresses or copied setup commands. If you choose to send feedback or a screenshot, check what you include before sending it.
Choose an appropriate first source
Start with a document you are allowed to use with the selected processing services. Try a known question and a question the file cannot answer. A useful first check verifies both a supported answer and an honest statement that information is missing.
The knowledge base starter kit includes fictional examples and a source register. Use it to practice the workflow before introducing internal material. For questions about processing, access or deletion, contact [email protected].