Kogbi RU

Language models that run inside your company

We set up language models, search over your documents and fine-tuning on hardware you control. No question and no document goes outside.

Will the model fit in your memory?

model size, billions of weights
bits per weight
your memory, GB

Fits in 24 GB. It will run fast.

Needs about 21 GB: a 19 GB model file plus about 2 GB for context.

A rough estimate. Memory is the main limit: video memory on a PC, shared memory on a Mac.

Why run a model yourself

Confidentiality
Neither the question nor the document leaves your network.
Independence
No outside service, subscription or payment card to depend on.
Stability
Nobody updates or switches off the model without you.
Works offline
The model answers with no internet connection.
Fitted to you
Your own instructions and fine-tuning.

When local is the wrong choice

A local model is noticeably weaker than the best cloud models. We say so before a project starts, not after.

run locally

  • Sensitive data
  • High-volume, repetitive work

keep in the cloud for now

  • Complex reasoning
  • Long tasks
  • Agents

The setup that works is a hybrid. We help you draw the line.

What we do

Local deployment
We choose a model for your tasks and your hardware, quantize it and set up inference on a server, a workstation or a laptop.
Answers from your documents (RAG)
The model answers from your documents and shows its sources. Update a document and the answer changes straight away.
Embedding models
This is what Kogbi builds as a company. The embedding model decides which fragments the search finds, so RAG quality starts here.
Fine-tuning (LoRA)
For style, tone, a strict output format or one narrow task. Not for loading knowledge: that is what RAG is for.
Hardware sizing
Don't buy hardware in advance. We test on what you have, then work out what you need: memory, quantization, model architecture.
Team training
How models work, with no magic: tokens, weights, context, and the places where they get things wrong.

How we work

  1. Start with what you have

    A pilot on your existing computers shows what a local model can do for your task and what it can't.

  2. Put the documents in order

    Three versions of a price list give three different answers. We clean up the sources before indexing them.

  3. Agree on control questions

    A fixed set of questions with known answers. Without it nobody can tell whether a change made things better or worse.

  4. Prompt, then RAG, then LoRA

    In that order. Each step costs more than the one before, so we take it only when the previous one isn't enough.

  5. Deploy and hand over

    On your hardware, documented, so your team can run it without us.

RAG: the model answers from your documents

A model doesn't know your business, and your documents won't all fit in its context. RAG finds the relevant fragments and puts them in the prompt.

once
  1. Documents
  2. Fragments
  3. Index
on every question
  1. Question
  2. Search
  3. Model
  4. Answer with sources

Where RAG breaks

Messy documents
Three versions of a price list, three different answers.
Scans and tables
What was recognized badly will not be found.
Questions about the whole base
RAG searches. It doesn't count.
No control questions
You can't see whether it got better or worse.

RAG is a project about order in your documents.

What hardware it takes

Laptop
16 GB
Small models: drafts, sorting, data extraction.
Mac
32–128 GB
The simplest route: memory is shared between processor and graphics.
PC with NVIDIA
16–32 GB of video memory
Faster, but the model is smaller.
Server with GPUs
as needed
Large models, many users at once.

A comfortable speed starts at about 20 tokens per second.

Team

Kogbi GmbH is a startup that builds embedding models for RAG.

Ivan Pleshkov

Ivan Pleshkov

CEO and co-founder

Qdrant
Four years as a core developer of the vector database
10+ years
in engineering

Tell us what the model should do

Describe the task, the documents and the hardware you have. We'll tell you what is realistic to run locally and what isn't.

Email hello@kogbi.com