LLM Orchestration

LLM orchestration of agents allows for much more nuanced and analytical output than you can get from a single LLM.

This is the main Khoraxis Window. On the left you can instantiate an LLM.

Khoraxis Workspace

Like so…

A single LLM instance

You can designate which model to use (from both cloud and local providers) as well as fallback options. You can adjust the weight its outputs will contribute overall. Then there are standard options: Initialize with instructions, add local knowledge, view the chat history, and classify by Role Type (this allows for both default parameters and custom Role configurations). You can even chat locally with the model.

Nothing special.

Let’s add a couple more LLMs.

Three instanced LLMs

Now we have a network. The arrows designate allowed signal flow and can be changed to one direction only. The midpoints gather outputs and can weight them.

This would be a useful setup for a code review. LLM 21 could focus on the code that’s present, while LLM 22 could focus on what the code is missing. They could then pass their analyses into LLM 23 for a final synthesis.

Orchestration is handled through Task Blocks like the following:

A Khoraxis Task Block

We are feeding in a text and asking for a ClaimSet output (this will give us a formatted list of claim pairings). In this instance we are analyzing company SEC 10-K Risk Disclosures. Each block can tag a single or multiple LLMs that can be synced to a single step in the orchestration chain.

Task Blocks are strung together into an overall sequence.

A sequence of Task Blocks

Each Task Block has its own inputs and outputs, its own set of tasked LLMs, its own function types that it can process.

So in our demo example we created an AI Review Board that compares official SEC 10-K risk disclosures against CFPB consumer complaint narratives, detects sentiment and trust erosion, identifies under-disclosed operational risks, and produces an auditable executive risk brief.

It looks like this:

Khoraxis Network

Every LLM can be set up as its own expert agent and synchronized within a network of expert agents.

There are a number of preset Workflow Templates that cover different use-cases.

That’s all for now. More to come.