Ask like you would ask a colleague
The answer arrives as it is written, with tables, lists and quotations where they help. Follow-up questions keep the thread.
HR
FIN
SCM
CRM
PM
DB
ERP
TATHVEN AGENTIC ERP
Tathven AI Workspaces · document management
Every document you have.
Every document you have.
Every answer inside them.
Tathven AI Workspaces stores, versions and secures your documents. Folders, permissions, versions and audit behave the way people expect. On top of that, an assistant reads the documents you are allowed to see, answers in plain language and cites each source. It can run on your own infrastructure.
I think it’s in one of these folders” is an expensive sentence.
Permission first
Checked before any text reaches a model, not afterwards on the citations.
Reads whole documents
Then says how much of your archive the answer was based on.
AI is a right you grant
Off by default, per person. Without it, people see a plain DMS.
Metered in the open
Tokens, cost, time and model for every question.
Ask, in plain language
Ask a question.
Ask a question.
Get an answer, with its sources.
Answers are built from the full text of the documents that matter, with every claim citing the document it came from, so you can check it. A file server can tell you a file exists. This can tell you what is in it.
Every claim cites its document
Inline citations name the source, the sources sit underneath as chips, and each chip opens the document it names.
Pin one document
Type @ and a filename to ask about that document alone.
Summarise a whole folder
A reading of the entire folder and everything beneath it, built in the background so asking is usually instant.
It says how much it read
When a document was too long to read in one pass, or a folder held more than was read, the answer says so rather than sounding complete.
Translate a document
Hindi and Tamil today. Translating needs the Download right, because a translation is the content in copyable form.
The honest part: every folder summary ends by saying what it was built from.
Find by meaning, or by name
Three ways to look, and they are honestly different
Match a name, find by meaning, or ask a question. The product never pretends one is the other, and only one of them spends a model call.
Names
Filename match as you type, with wildcards like *.zip.
Deep
Documents ranked by what they are about, and by keyword, with the passage that matched and its score.
No model call
Ask
An answer, built from reading the documents end to end, with citations.
Uses a model
1Upload
Into any folder
2Consent
An administrator switches indexing on, folder by folder
3Index
Searchable in about a minute
4Ask
Only over what you may open
5Cited answer
Every claim names its document
The document plane
Folders that behave
Folders that behave
the way people expect
Permissions inherit from the parent and break where you need them to, every version is kept, and every deletion is recoverable. The familiar gestures, with a record of who did what.
Six rights, inherited
Read, Download, Modify, Delete, Create and Manage permissions, for people or groups, on folders or single documents. Rights flow down the tree; break inheritance on any item and it keeps a snapshot, then diverges. Grant-only: no deny rules to reason about.
“Read only” means read only
Read without Download renders the document in the browser and the file never leaves. A watermark with your wording and logo is drawn on every such view.
Vaults
Password-protected folders, sealed by default and re-sealed after an idle timeout. A vault’s contents are never indexed, because an indexed vault would not be a vault.
Versions that cannot be rewritten
Every version is kept. Restoring an old one writes a new version rather than rewriting history.
Who opened what, and when
Views, downloads, edits, permission changes, restores and every administrator bypass are written to the audit trail.
Nothing lost by accident
Deletions wait in Trash until it is emptied, and storage tells you honestly what it counts: current documents, superseded versions and trash.
Read and edit in the browser
It opens what
It opens what
you already have
Documents preview in place: no download, no desktop application, nothing to install. The list sits on the left, the document on the right, and its version history underneath.
Documents
PDF · Word · Excel · PowerPoint · Markdown · HTML · plain text · RTF
Books
EPUB with its own table of contents · CBZ page reader
Data
JSON and XML as a collapsible tree · CSV as a sheet
Diagrams
Mermaid · Graphviz DOT · draw.io
Media
Images, audio and video, played in place
Archives
ZIP · RAR · 7z · TAR · GZ, extracted in place or into a new folder
Edit online, one version per session
A full office suite runs inside the workspace. Autosaves land on a working copy and one version is sealed per editing session, so the history stays readable. Close the tab and the work is sealed anyway.
Nobody overwrites anybody
The document is checked out while you have it open. If someone uploaded a new version meanwhile, you are told, and your work is kept.
The editor follows the rights
Print and copy need Download. Export, Save a copy, plugins and macros are off for everyone, administrators included.
Editing in the browser is provided by ONLYOFFICE Docs, a trademark of Ascensio System SIA. It runs unmodified as its own service, and Tathven talks to it over WOPI. Tathven is not affiliated with or endorsed by Ascensio System SIA. The complete corresponding source, and the exact change, are published at github.com/tathven-enterprise/ONLYOFFICE.
Know what it costs · your models, your choice
Every question is metered,
Every question is metered,
and you can see it
Tokens, cost, duration and model for every query, so nobody has to guess what the assistant is spending. Run it entirely on your own machines, or use a frontier model with your own key.
Today, 7 days and 30 days, with a per-model breakdown.
A per-query ledger: who asked, which mode and model, tokens in and out, model calls, seconds and cost.
Local models through Ollama: no document text leaves the network.
Or a hosted model (Anthropic, OpenAI) with your own key. The picker labels it “sends content off-site” and warns again beneath.
Keys encrypted at rest, shown once, then only as their last four characters. Test connection makes a real call, so an administrator finds a bad key before a user does.
Six commitments
An assistant you can
An assistant you can
trust with the archive
It can only ever read what the person asking is allowed to read. These are the product’s own words, from its About page.
1. Your documents stay on your infrastructure
The search index and the models that read your documents run inside your own network. Document contents are not sent to a third party unless you configure a hosted model yourself.
2. Indexing is switched on folder by folder
Reading a document to index it is treated as an access decision, not a setting. It is off until an administrator enables it, and every change is recorded with who made it.
3. The assistant can only read what you can
Permissions are checked before any document reaches a model, not afterwards on the citations. If you cannot open a document, the assistant cannot quote it either.
4. Password-protected folders are never indexed
Content in a vault is excluded outright, because an indexed vault would not be a vault.
5. Answers say how much they are based on
When a document was too long to read whole, or a folder summary covered only part of what is filed, the answer says so rather than sounding complete.
6. Every content access is audited
“Was my document used to answer somebody’s question?” is a question this system can answer.
What leaves the network? Models, embeddings and text extraction all run inside it. The single exception is opening a raw .drawio file, which loads draw.io’s viewer from the internet; the file’s contents never leave the browser, and an air-gapped site can point that at its own copy.
Not trained on your documents. The assistant reads them at the moment you ask. Nothing is fine-tuned, and no document becomes training data.
Read-only is a deterrent, not a guarantee. The watermark, the audit entry and never exposing a raw address are the controls that hold; a determined viewer can still capture what their own screen shows.
Where it runs
One product,
One product,
three editions
One codebase. An edition is a configuration, never a fork, so every feature arrives in all three.
Personal
- Runs on
- Your own laptop
- Documents live
- On your disk
- People
- One owner
- AI
- All on, with your own model keys
On-premise
- Runs on
- Your servers
- Documents live
- On your servers
- People
- The full rights model, local accounts
- AI
- Granted per person, with your own model keys
Cloud
- Runs on
- Hosted by Tathven, on your organisation’s own address
- Documents live
- In a database, storage bucket and search index of your organisation’s own
- People
- Sign in with Google, or email and password; optionally only your Google Workspace domain
- AI
- Granted per person within the workspace
~40 sto answer across a folder, reading every matching document in full
a few ¢US cents for a typical folder-wide question, measured in the product
~1 minfrom upload to searchable: the indexer sweeps every 60 seconds
2,100+documents in 900+ folders, indexed and answerable in one archive in production
Built on PostgreSQL 17, .NET 10, Angular and Docker. Screenshots are from a working workspace; one person’s name has been changed.
Request a demo
See Tathven AI Workspaces running on
See Tathven AI Workspaces running on
your own documents
Tell us what you run today. We will show Tathven AI Workspaces working on a sample of your own documents, then send a written scope and quotation — not a brochure.