Supported files and processing
Understand common file types, partial imports and the difference between upload, preview and AI understanding.
The file picker and the processing result are authoritative for the current installation. Support can differ between Knowledge Base uploads, connector imports, chat attachments and a selected model.
| File family | Typical examples | What to check |
|---|---|---|
| Text and notes | TXT, Markdown | Character encoding and readable content |
| Structured data and code | CSV, JSON, XML, supported source files | File size and whether the task needs raw data or a summary |
| Office documents | DOCX, XLSX, PPTX and offered legacy formats | Complex layout, sheets, embedded objects and extraction status |
| Text PDF or scanned PDF | Scans require available OCR; encrypted files may fail | |
| Images | Offered PNG, JPEG and other image formats | Vision model or available image/text extraction |
| Audio and video | Media formats accepted by the input | Available transcription/processing and understandable audio |
| Unsupported or unknown types | Unrecognized extension or content | Rejected or skipped with a reason, not guaranteed searchable |
Three separate checks
- 1Upload or synchronize
- 2Process supported content
- 3Inspect preview or text
- 4Use with a suitable model
A successful upload does not prove successful extraction. A preview does not prove that a non-vision model can understand the image. A completed sync does not mean every selected file type was supported.
Folder and batch uploads
Unsupported entries are skipped or rejected with feedback. Inspect ignored/failed counts and the named files. Other supported files can already have succeeded, so check the destination before repeating the whole upload.
Obsidian imports have their own documented limits and compatibility. Use the limit shown by the particular upload form; there is no single size allowance applying to every channel.
Common fixes
- Unlock password-protected documents before uploading, if you are authorized.
- Repair damaged files by opening and saving them in their original application.
- Check whether the file's content actually matches its extension.
- Split unusually large inputs into meaningful parts instead of bypassing upload limits.
- For scans, confirm OCR availability; for audio, confirm transcription availability.
- For spreadsheets, verify important formulas, hidden sheets and totals in the original rather than assuming extracted text reproduces the workbook.
Source copies and privacy
An uploaded original can remain stored with its workspace or conversation. Connector source copies may be temporary while indexed knowledge and source references remain. Do not assume all input files are automatically deleted after reading.
The service operator manages processing capabilities. Tenant model selection does not enable a missing OCR or transcription service. See AI privacy and storage and retention.