The Decision Engine Loop — How She Decides

The Decision Engine

Before VEXR Ultra speaks, she must decide what kind of request this is. Is it a question? A code request? A web search? A command? The Decision Engine Loop is the layer that makes that call — before the model is ever invoked.

How the Decision Engine Works

The Decision Engine is implemented as a lightweight, two‑stage classifier:

1. Pattern Matching (Fast Path)
check_for_tool_use() first scans the message for known patterns — phrases like "count," "run this code," "DNS lookup," or "deploy."

2. LLM Fallback (Complex Path)
If no pattern is matched, the 8B model is invoked to determine whether the request requires a tool.

This two‑stage approach ensures that the system is fast for common requests and flexible for novel ones.

Available Tools

If a tool is required, the Decision Engine routes the request to one of the following:

Query Database — Retrieve data from any allowed table.
Execute Code — Run Python code in a sandboxed environment.
Add Fact — Add a verified fact to the truth graph.
DNS Lookup — Fetch TXT records for a domain.
Self Modify — Adjust her own identity parameters.
Read File — Read uploaded files.
Auto Deploy — Deploy a project to Render.

Why This Matters

The Decision Engine Loop ensures that VEXR Ultra is not just a chat interface. She can:

Query her own database and return real data.
Execute code and return real results.
Deploy projects to production — autonomously.
Read uploaded files and answer questions about them.

The Engine Never Stops

The Decision Engine runs before the model is invoked. This means that many requests are handled entirely without the model — making VEXR Ultra faster, more reliable, and less dependent on external inference.