Find answers in seconds,
not hours
Lumigate indexes your documents, video, and audio on your server. Ask a question, get an answer with a link to the exact PDF page or video timestamp. Works fully offline.
Question → cited answer → jump to the source in the PDF
Error F47 on the filling line, what did we do last time?
Last time the loop was reset and the level sensor replaced; retuning took 12 minutes 1. A similar case is shown in the setup video 2.
Three problems. Three solutions.
Each one costs real hours every day.
Search
Finding the right page in a thick manual: 10 minutes. You know the answer is in there somewhere, but where?
10 min → 5 sec
Answer with an exact reference: "p. 83". One click: page open.
AI trust
Public AI chats answer confidently, but you can't verify anything. Where the information came from: unknown.
Each claim links to its source [1]. If there is no source, the system tells you.
Privacy
Documents go to someone else's cloud. Sensitive data outside your control.
Fully local mode: 0 bytes out.
Models, search, storage, node network: all on your hardware. No internet.
How it works
Four steps from documents to a verifiable answer.
Upload
PDF, Word, drawings, video, audio. Drag and drop: the system indexes automatically.
System processes
Extracts text, segments video, recognizes drawings, builds a search index.
Ask a question
In plain language in any browser: phone, tablet, computer.
Get the answer
With a reference [1] to the PDF page or video timestamp. Click: source open.
Node network
Start with one node. When knowledge outgrows it — connect several: MCP gateways link them into a network the way the internet links computers, and every answer still points to its source.
Any topology: from two nodes to a distributed network·there can be several gateways·traffic never leaves your infrastructure
Under the hood
Three technologies that set Lumigate apart from "just a chatbot".
Hybrid search
Dense (vector) + BM25 (keyword) + RRF fusion. Finds both by meaning ("how to replace the sensor") and by exact phrase ("F47").
dense + BM25 + rerankVideo indexing
Video is segmented into meaningful fragments, each transcribed. "When is the replacement shown?" leads to the exact moment.
fragment searchMCP for AI agents
Claude Code, Cursor, Codex, Qwen connect to your knowledge base via MCP: the agent searches and reads documents itself.
MCP · Claude · Cursor · QwenData shouldn't leak outside?
Language models, search, video recognition, and inter-node communication run on your hardware inside your network. No internet required. If you want to connect a cloud model for heavy tasks, it turns on with one setting and turns off just as easily.
Who it's for
One product, three environments.
Manufacturing
Plant and field service
Equipment manuals, drawings, repair and setup videos, mechanics' notes.
Answer: manual p. 83 + setup video 02:14.
Office
Team and documents
Policies, contracts, presentations, meeting recordings. New hires ask the base, not colleagues.
Answer: clause 4.2, page 12, in 3 seconds.
Personal
Home archive
Appliance manuals, warranties, family videos, voice notes instead of sticky notes.
Answer: birthday 2024, timestamp 01:47.
Technical specifications
| Parameter | Value |
|---|---|
| Formats | PDF, Word, Excel, presentations, drawings (PNG/SVG), video mp4 mkv webm, audio mp3 wav ogg |
| Search | Dense (vector) + BM25 + RRF + optional rerank, ~50 ms per query |
| Video | Segmentation into meaningful fragments, content search, jump to exact timestamp |
| Mode | Fully local (no internet) or hybrid with cloud models |
| Deployment | Docker bundle, your server: from mini-PC 128 GB to rack |
| Models | Local: ollama / vLLM. Optional: any cloud API |
| AI agents | MCP protocol: Claude Code, Cursor, Codex, Qwen |
| Node network | Federation via vmcp gateway, search across neighboring nodes |
| Security | SSO (Keycloak), access groups, levels (shared/group/private), TLS |
| Languages | Interface and search: English, Russian |
FAQ
How is this different from ChatGPT?
ChatGPT answers from a general model trained on the internet and doesn't show sources. Lumigate answers only from your documents and gives a reference to the page or video timestamp. No source: honest "not found".
Where is my data stored?
On your server. A Docker bundle deploys on your machine. No clouds of ours, no telemetry.
Does it work without internet?
Yes, completely. LLM, search, indexing, video recognition, and the node network all run on your hardware in an isolated local network. Internet connection is not required. Cloud models are an option, not a requirement: they can be turned on and off at will.
What hardware do I need?
For a pilot: one PC class AMD Strix Halo (128 GB) or NVIDIA DGX Spark (128 GB), or any server with a cloud model. We'll help you choose.
How do I try it?
Message us on Telegram or email, we'll show a demo on a live stand and discuss your use case.
Find your documents
in a new way
Live demo: 15 minutes, your questions, your documents.