Documents change. Answers follow.
fine-tune: retrain the modelrag: re-index the changed file
Update a manual and answers use the new version as soon as it’s re-indexed.
RAGRetrieval-augmented generation · on your own server
Soshyant Expert is a RAG engine that runs on your own server. It retrieves the exact passages from your manuals, policies and procedures — then generates an answer grounded in them, with every claim cited.
- Retrieval
- Hybrid
- Embeddings
- Local · 384-d
- Citations
- Enforced
- Cloud uplink
- None
- Scans
- English OCR
01 // Live console
Ask. Retrieve.
Answer. Cite.
Every answer is generated from passages pulled out of your own documents — and it shows you exactly which ones. Pick a session to replay it.
you ›What’s the torque spec for the M12 bolts on the pump housing?
Tighten them to 60 N·m in a star pattern, in two passes — 30 N·m first, then 60 N·m. Re-check the torque after the first heat cycle.1
1Pump Manual rev. 4§7.2 Assembly · version 4
Torque the M12 housing bolts to 60 N·m in a star pattern, in two passes (30 N·m, then 60 N·m). Re-torque after the first heat cycle.
02 // Signal loss
19%
of the average workweek is spent searching for and gathering internal information.
Source: McKinsey Global Institute, “The Social Economy” (2012)Your knowledge is buried alive.
- ERR_01 · repeat_query
Experts answer the same questions, again and again
Senior staff spend hours repeating what the manual already says.
- ERR_02 · keyword_miss
Keyword search misses the answer
It fails on procedures, part numbers and questions phrased differently from the document.
- ERR_03 · knowledge_drain
Expertise walks out the door
When people leave, their know-how leaves with them. The documents stay — but stop being usable.
03 // RAG, decoded
Retrieval-augmented
generation.
A language model on its own answers from memory — and memory can be wrong. RAG changes the order: first retrieve the passages that matter from your documents, then augment the prompt with that evidence, and only then generate the answer. Run the pipeline yourself:
- 01Embedquestion → vector
- 02Retrievehybrid search
- 03Filteraccess control
- 04Augmentbuild the prompt
- 05Generatecited answer
// query
›What torque for the M12 pump housing bolts?
// embedding 384-d · multilingual · local
[ … ]
// knowledge base
lexicalsemanticaccess denied
// context window 0 / 8,192 tokens
// answer
Waiting for pipeline…
Why RAG — not fine-tuning?
Knowledge is data, not weights.
fine-tune: baked into weightsrag: versioned records you control
Your content stays in a database on your server — versioned and backed up like any business record.
Citations need sources.
fine-tune: can’t point to a sourcerag: every claim → a passage
Answers point at real passages — document, section and version — so people can check them.
04 // System modules
Built like infrastructure.
Feels like an expert.
mod.01 // cite
Answers you can check
Every claim links to the passage it came from — document, section and version. Click a citation and read the evidence yourself.
mod.02 // limits
Says when it doesn’t know
No evidence, no answer. It asks a clarifying question or says the documents don’t cover it.
mod.03 // hybrid
Hybrid search
Keywords catch part numbers and exact terms. Vectors catch meaning. Soshyant fuses both.
mod.04 // private
Private by default
The language model, the embedding model, the index and every document live on your server. In local mode: no CDN calls, no automatic downloads, no cloud.
mod.05 // access
Permission-aware retrieval
Access groups decide who can use each document. Restricted passages are removed on the server before ranking — the model never sees what the user can’t.
| doc | field | sales | eng |
|---|---|---|---|
| Safety proc. | |||
| Pricing 2026 | |||
| Design specs |
mod.06 // lang
English & فارسی
Switch interface and answer language instantly, with a full right-to-left layout.
mod.07 // watch
Watched folders
New and changed files are detected and wait for an admin’s approval before anything is published.
scan ./manuals · 2 changed · awaiting approval
mod.08 // ocr
Reads scanned English pages
Scanned PDFs and phone photos are read by a local OCR model, then indexed and cited like any other page — as “page 4 (OCR)”. Pages it can’t read reliably, and Persian scans, are flagged for an admin instead of guessed.
mod.09 // review
Answer review
Triage unanswered questions and user feedback, and note which document needs fixing.
queue · insufficient · clarify · feedback
mod.10 // models
Bring your model
Use the bundled local model, any GGUF model on your server, or an OpenAI-compatible API — only after an admin explicitly approves sending context.
mod.11 // plan
Hardware planner
Estimate GPUs, RAM and CPU for your number of users — before you buy anything.
users → gpu · ram · cpu
mod.12 // backup
Backup & restore
Back up documents, chats, accounts and settings. Restores are validated before they replace anything.
backup.zip · verified
05 // Perimeter
Nothing leaves
the building.
For manufacturers, energy, healthcare and defense suppliers, keeping documents in-house is a requirement — not a feature. Flip the switch to see the difference.
- Your documentsStay on your own server
- The modelRuns on your hardware
- CostNo per-question API fees (local mode)
- Answers fromOnly your approved documents
- ProofDocument, section and version cited
security --audit
- [ OK ]No default accounts — the first admin is created with a one-time code from the server
- [ OK ]Hashed passwords, server-side sessions, CSRF protection on every change
- [ OK ]Conversations isolated per user — identity always comes from the session
- [ OK ]TLS on your intranet with your own certificate and allowed host names
- [ OK ]Permissions enforced before search, model context, downloads and history
- [ OK ]Question and document text kept out of HTTP logs by default
- [ OK ]Optional API keys stay server-side, encrypted with Windows DPAPI
- [ OK ]Documents and evidence published in one atomic transaction
- › audit complete · 8/8 checks passed
06 // Deployment targets
Runs where your team works.
Windows desktop launcher
Double-click to start. It runs quietly in the system tray with clean start, stop and server logs — no command window.
system trayoffline install
Your internal network
Every browser on your intranet connects over HTTPS. The language model itself only ever listens on the server.
https / tlsbrowser-based
Android companion app
Answers on the shop floor. A native app that connects over your Wi-Fi with your private certificate — and blocks screenshots.
native androidwi-fi + https
07 // Establish uplink
Book a live demo.
Pick a slot. We’ll run the pipeline live and talk about what a pilot would look like for your team.
- 01
Live pipeline run
Ask questions, open citations, and watch it refuse when the evidence isn’t there.
- 02
The admin console
Documents, watched folders, access groups and answer review.
- 03
Your deployment
Users, languages, hardware sizing and data boundaries.
- 04
A pilot plan
How to measure answer quality on your own questions.
Schedule your demogoogle calendar · secure booking
new tab
08 // Query log
Questions, answered.
Something else on your mind? Open a channel — we’re happy to go deep.
Q.01What does “RAG” mean here?
A ›
Retrieval-augmented generation. When someone asks a question, Soshyant first searches your indexed documents (keyword and semantic), keeps only the passages that person is allowed to see, and hands them to the language model with instructions to answer only from that evidence and cite it. If the passages don’t support an answer, it says so.
Q.02Does it train the model on our documents?
A ›
No. Documents are indexed for retrieval; the model’s weights are never changed. When you update a document, answers use the new version as soon as it’s re-indexed — no retraining.
Q.03Does it need an internet connection?
A ›
Not in local mode. The language model, the embedding model, the OCR model and the interface files all load from the server’s own disk — no CDN, web fonts or automatic downloads. If an administrator chooses to connect a cloud model, the server then needs internet and the question plus retrieved passages go to that provider — only after explicit approval.
Q.04Which documents can it read?
A ›
PDF, Word (DOCX), Excel (XLSX), plain text, Markdown and HTML, plus scanned images (PNG, JPG, TIFF). Scanned English pages are read by a local OCR model; Persian scans need a text version for now. Upload files one by one, import a folder, or watch a folder for changes.
Q.05Which languages are supported?
A ›
The interface and answers are available in English and Persian (فارسی), with a full right-to-left layout. We recommend validating answer quality on real questions from your own domain and language during a pilot.
Q.06Can different teams see different documents?
A ›
Yes. Administrators create access groups and choose which groups can use each document. Restricted passages are filtered on the server before ranking and before anything reaches the model. Confidential files can be imported as “administrators only” from the first second.
Q.07What hardware do we need?
A ›
A single Windows server or workstation. A GPU is recommended for responsive answers; CPU-only operation is possible. The built-in hardware planner estimates GPUs, RAM and CPU cores for your number of users — and we’ll size it with you during the demo.
Q.08Is there a mobile app?
A ›
Yes — a native Android app that connects to your own server over Wi-Fi using HTTPS and your private certificate. The model never runs on the phone, passwords aren’t stored on it, and screenshots are blocked.
Q.09Where is our data stored?
A ›
Documents, conversations and accounts live in a local database on your server, and backups are written there too. You decide how that storage is protected and retained under your own policies.
09 // Open a channel
The companies that wrote the manuals deserve an expert that never forgets.
Tell us about your documents and your team — we’ll show you what a RAG expert running on your own hardware can do with them.
operator idsoshyant // expert
Saeid SalemiFounder and CTPO
- emailSaeed@ahouda.com
- phone+98 999 993 4659
- linkedinlinkedin.com/in/saeid-salemi
- websitesoshyantexpert.site
clearance: public