This page is machine-translated from French. Read the French original.
File formats
Text and documentation
| Scale | Treatment |
|---|---|
.txt |
Direct reading |
.md |
Direct reading, structure preserved |
.csv |
Direct reading, line by line |
.json |
Direct reading |
.html |
Text extracted, tags removed |
.org, .adoc, .rst |
Direct reading |
Office Automation
| Scale | Treatment |
|---|---|
.pdf |
Text extraction; character recognition if the PDF is an image |
.docx |
Extracting the text and structure |
.pptx |
Extracting text from slides |
.xlsx |
Extracting cells, sheet by sheet |
.odt, .odp |
OpenDocument, text and presentation formats |
The format
.docWord (previously 2007) is not supported. Convert to.docxbeforehand.
Messaging and books
| Scale | Treatment |
|---|---|
.mbox |
Email archive; each message becomes an indexable item |
.epub |
Digital book, text extracted chapter by chapter |
Images
| Extensions | Treatment |
|---|---|
.png, .jpg, .webp |
Character recognition — the text in the image is extracted |
The image itself is not indexed; only the text it contains is. Quality depends directly on sharpness and resolution.
Audio and video
| Extensions | Treatment |
|---|---|
.mp3, .wav, .m4a, .ogg, .opus, .webm |
Automatic transcription |
.mp4, .mpeg |
Extracting the audio track and then transcribing it |
The integrated transcription is running locally, without a key or network output. Processing time increases with recording time.
Remote sources
| Source | What is imported |
|---|---|
| Web page | Main content, menus and advertisements kept out of the way as much as possible |
| Full site | Recursive exploration, configurable depth and number of pages |
| Code repositories (GitHub, GitLab) | Files in a repository, filterable by branch and by extension |
| Confluence | Pages of a space |
| Corporate Wikis | Pages and tree structure |
| Document management (Paperless-ngx) | Documents and metadata |
| Note chests (Obsidian) | Notes and links |
| YouTube videos | Transcription |
This leads to poor results
Low-quality scanned PDFs. Character recognition applies, but a tilted, smudged, or low-resolution scan produces faulty text — and therefore faulty responses.
Complex spreadsheets. Merged cells, formulas, pivot tables, and very wide worksheets do not linearize well. For large tabular data, the agent's SQL skills provide much better results. → SKILLS
Very rich layouts. The text is extracted; the information conveyed solely by the visual arrangement — a diagram, a graphic table — is lost.
Poor quality audio recordings. Distant voices, overlapping speech, and background noise severely degrade the transcription.
Very large documents. A document of several thousand pages produces a large number of fragments and dilutes the search. Divide it into coherent units whenever possible.
Check what has been extracted
The exact list of formats accepted by your The instance is available in the import interface and via the API. → API developer
After importing a difficult document, ask a question to which you know the answer and look at the quoted excerpt: this is the fastest way to know if the extraction was successful.