Skip to content
Service status

Type to search

1 articles · 4 min read

Virtual machine

Every assistant has its own sealed-off computer for the conversation, and runs programs on it instead of working everything out in its context.

Every assistant in Intric has access to its own virtual machine (VM) for the length of a conversation. A virtual machine is a computer that runs entirely in software: it has its own operating system, its own files and its own programs, but no physical hardware of its own. Think of it as a private workstation that belongs to that one conversation and nothing else.

The assistant decides when to use it. You don’t switch it on per assistant, and you don’t need to ask for it, although you can (“work this out with a script” is a perfectly good instruction).

Without a computer, everything an assistant does happens inside the model’s context window: its short-term memory, everything it can “see” at once. Every document, instruction and intermediate result has to fit in there, and every new skill — reading a CSV file, opening a spreadsheet — has to be added as a separate tool, whose description takes up room too. Large tasks fill the context quickly. That costs more, and models answer worse when their context is crowded.

With a VM, the assistant can recognise that a problem is best solved by a program, write that program and run it, instead of reasoning its way to the answer by reading:

  • The work moves out of the conversation. Files and programs work together on the VM, and only the result comes back. The context stays clear.
  • One general ability replaces many narrow ones. Instead of dozens of separate tools, the assistant has one: using a computer. Today’s language models have been trained on a great deal of material about working in a Linux terminal, so they already know how.
  • Exact answers instead of estimates. Arithmetic, counting, dates, sorting and comparing are computed, not guessed.
  • It works on real files. The assistant can open, analyse, edit and create CSV, Excel, Word, PowerPoint and PDF files, and hand the result back to you as a file you can download.

You ask: “Compare this year’s annual report with last year’s and list every key figure that changed by more than 10%.”

Without a VM, the assistant has to load both reports into its context and find the differences by reading. That is slow, expensive, and with hundreds of figures it is easy to miss one.

With a VM, the assistant extracts the tables from both files and compares them with a short program. You get an exact list of every difference, and you can download it as an Excel file.

You upload your organisation’s Word template for a decision memo, together with a case file and your notes, and ask: “Fill in the template for this case and give me a PDF I can send.”

Without a VM, the assistant can write the text out in the chat for you to copy into the template yourself, or it can create the memo as an artefact: a document that opens beside the conversation, built on your template, that you can edit and download. An artefact is a good fit for a short document you want to read and adjust as it takes shape. But the model writes every word itself, so the whole document has to pass through the context. A change you ask for doesn’t necessarily mean the model rewrites the entire artefact, but everything it works on still goes through the context, and for a long document, or a template with many fields, that uses a lot of tokens.

With a VM, the assistant opens your Word file itself, writes each section into the right place and keeps your headings, logo and layout as they are. Only the text it adds has to pass through the context, not the template around it. It then converts the result to PDF, and you get both the finished Word file and the PDF to download.

The machine has no internet connection, every conversation gets its own, and the assistant cannot change the files you share with it. For what is installed on the machine and how it is kept secure, see Technical details.

If you run Intric in the cloud, on a dedicated or a multi-tenant instance, the VM is already on for every assistant. There is nothing to switch on.

On on-prem and self-hosted installations it is off until it has been enabled. It needs a server with hardware virtualisation (KVM) and some extra capacity per concurrent machine; see Server and hardware.