Documents
Text and office documents
Native text is read directly; the parser reads layout and tables.
pdf · docx · doc · docm · odt · rtf · txt · rst · md · mdx · epub · html · xhtml · json
Turn documents, sites and wikis into knowledge your AI can cite
DocsGPT turns your files, web pages and wikis into a private knowledge base: 30+ formats including scanned PDFs, parsed, chunked and embedded so every answer cites its source. Upload, link or crawl, sync from your connectors, or add files straight from a chat.
upload · link · crawler · wiki · connectors · ocr · 30+ formats
Last updated:
Ways to add knowledge
Upload files, add a link, crawl a website or start a wiki. Files attached in a chat can join Knowledge too: Add to Knowledge makes them searchable in later conversations and by your agents.
Share any source with your team as Viewer or Editor.
upload · link · crawler · new wiki · add to knowledge · viewer · editor
Synced from your connectors
Connect Google Drive, SharePoint, Confluence, Amazon S3, GitHub or Linear once, and the content you pick syncs into Knowledge on a schedule. Synced files are chunked, searched and cited like any upload, marked with the service they came from.
See all connectors →- Google Drive
- SharePoint
- Amazon S3
- Confluence
- Linear
- GitHub
File formats (30+)
Thirty-plus formats are parsed on the DocsGPT backend into text, tables and Markdown before chunking. Audio is transcribed first; images and scans go through OCR.
ZIP archives are an accepted upload too — the archive is unpacked and every file inside is parsed.
Files attached in chat take a few more: tiff · tif · bmp · webp · adoc · asciidoc · vtt · xml
Default limits
- Per request
- 256 MB
- Per file
- 100 MB
- ZIP archive
- 250 MB uncompressed · 10,000 files
Defaults, not product ceilings: on a self-hosted deployment each of the three is an environment variable.
Data and slides
Spreadsheets and presentations
Sheets and slides are parsed cell by cell and slide by slide.
csv · xlsx · xls · xlsm · xlsb · ods · pptx · pptm · ppt · pps · ppsx · ppsm · pot · odp
Images
Images and scans
Image files and scanned pages go through OCR.
png · jpg · jpeg
Audio
Audio, transcribed on the backend
Recordings are transcribed before indexing; the transcript is the source.
mp3 · wav · m4a · ogg · webm
OCR for scanned PDFs and complex layouts
Scanned pages and images are read with OCR before chunking, so a contract signed on paper is as searchable as one typed in Word. The parser keeps layout and tables, and the result is clean Markdown.
Choose the OCR engine per deployment: tesseract, or DeepSeek-OCR run locally or hosted, with more engines through the optional Docling parser.
OCR applies to sources and to files attached in chat. For models that read images but not PDFs, DocsGPT sends the pages as images, so the model still sees them.
tesseract · deepseek-ocr · docling · sources + chat attachments
Ingest path for a scanned PDF
- 01Scanned PDFor image attachment
- 02Parserlayout · tables · text
- 03OCRtesseract · DeepSeek-OCR
- 04Text + tablesmarkdown
- 05Chunksstrategy per source
Text version
A scanned PDF (or image attachment) is parsed for layout, tables and text; OCR reads the scanned content with tesseract or DeepSeek-OCR; the result is text and tables in Markdown, which are chunked with the strategy set for that source.
Chunks you can see and edit
Five chunking strategies, chosen per source. Open any document to see its chunks with their token counts, and add, edit or delete chunks by hand. Test retrieval shows which chunks a question would find.
Strategies
- classic
- fixed-size chunks
- recursive
- split on structure, then size
- markdown
- split on headings
- parent_child
- small chunks retrieved, parent returned
- semantic
- split on meaning
token counts · add · edit · delete · test retrieval
Knowledge graphs from your documents
Turn on GraphRAG for a source and DocsGPT reads each chunk with an LLM and extracts the entities in it (people, companies, places, contracts) and how they relate. The result is a knowledge graph you can open next to the files.
Find an entity, see everything it connects to, and read a summary of what the documents say about it, with the number of chunks that mention it. Agents use the same graph to answer questions that span many documents, such as which customers depend on one carrier.
GraphRAG can be switched on when a source is added or later, and needs pgvector. How retrieval walks the graph is on search and retrieval.
graphrag · entities · relationships · mentioned in n chunks
Wiki sources your agents can edit
A Wiki source is a set of Markdown pages that agents can create, edit and rename. Every change carries version history and a provenance stamp, either human or agent, and the source re-embeds automatically so answers stay current.
Existing sources can be turned into wikis. The owner decides whether people using an agent through its API or website widget may edit the wiki too.
markdown · provenance: human | agent · version history · auto re-embed
How it runs privately
Parsing, OCR and embedding all run in the DocsGPT backend. On Managed, On-premises and Air-gapped deployments, files, OCR output and embeddings never leave your network.
Four deployment postures, one code base. Compare deployment options →
Inside your network
- upload · link · crawler · wiki
- parsers · ocr
- chunking · embeddings
- vector store · file storage
managed · on-premises · air-gapped
Cloud
Arc53 runs the backend; documents are parsed and embedded in Arc53's cloud.
Air-gapped
The same parsers, OCR and embeddings, with no external network. Uploads, wikis and sites on your own network work; cloud services such as Google Drive or SharePoint are not reachable.
Example
Government infrastructure, local inference
UK Department for Work and Pensions: an internal knowledge assistant on government infrastructure, with local inference.
Source: DWP
Next: sync from your tools · how answers are retrieved and cited · automate document processes · run it on-premises · compare with Microsoft 365 Copilot · pricing
Frequently asked questions
Does RAG work with PDF?
- Yes. Native-text PDFs are read directly, tables and layout are handled by the parser, and scanned pages are read with OCR before chunking and embedding.
Can AI read scanned PDFs?
- Yes. Scanned pages and images go through OCR, with tesseract or DeepSeek-OCR, and more engines through the optional Docling parser. The same applies to files attached in chat.
Can I add files from a chat to the knowledge base?
- Yes. Files attached to a conversation can be added to Knowledge in one step with Add to Knowledge, so they are searchable in later chats and by your agents, not only in the one conversation.
What file formats does DocsGPT support?
- Over thirty, in four groups. Documents: PDF, DOCX, DOC, DOCM, ODT, RTF, TXT, RST, MD/MDX, EPUB, HTML, XHTML, JSON. Spreadsheets and presentations: CSV, XLSX, XLS, XLSM, XLSB, ODS, PPTX, PPTM, PPT, PPS, PPSX, PPSM, POT, ODP. Images: PNG, JPG/JPEG. Audio: MP3, WAV, M4A, OGG, WebM, transcribed on the backend. ZIP archives are an accepted upload as well: they are unpacked and every file inside is parsed. Files attached in chat take four more image types (TIFF/TIF, BMP, WebP) and AsciiDoc (ADOC/ASCIIDOC), VTT and XML. The size figures — 256 MB per request, 100 MB per file, 250 MB uncompressed or 10,000 files per ZIP — are the defaults, and each is an environment variable on self-hosted.
Can the AI keep documentation up to date itself?
- Yes, with Wiki sources: Markdown pages that agents can create, edit and rename, with version history, a provenance stamp on every change (human or agent) and automatic re-embedding.
See DocsGPT on your documents
A 30-minute demo on your use case, or a free on-premises review with an engineer.