> ## Documentation Index
> Fetch the complete documentation index at: https://docs.drdroid.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Chaos Minimisation Framework (CMF)

> [WIP document]

If you're an engineer that has worked with building LLM & related applications, you know that the biggest challenge is to make it reliable & consistent.

At Dr. Droid, we are in the business of enabling engineers get a first responder agent for every issue so that the automation can assist in saving time & debugging issues faster in production.

This means:

* If an engineer asks "How's the health of service X? or when did service X get last deployed?", the answers needs to be (a) accurate (b) data backed (c) insightful (d) contextual.
* If an engineer asks "Run an analysis of why my user isn't able to do action A1", the agent needs to understand what is a "user" in context of the company, what does A1 mean and how it can check it.

LLMs come with their set of challenges, including but not limited to hallucinations, token limit, non-trivial data requirements for high-quality fine-tuning, prompt injections, misuse and lack of expertise on getting value from large volume of structured data.

This is where our CMF comes into picture. At Dr. Droid, we've solved for a lot of these challenges by building an agentic framework that's designed around minimising the chaos. Here's a quick glimpse of how the framework tackles these problems:

1. **Runbook automation framework:** Under the hoods, Dr. Droid deeply leverages the capabilities of [Playbooks](https://github.com/DrDroidLab/playbooks) -- a runbook automation framework.

2. **Contextual data access:**

   1. [Tools](https://docs.drdroid.io/docs/tools)
   2. [Memory](https://docs.drdroid.io/docs/memory)
   3. [Catalog](https://docs.drdroid.io/docs/catalog)
   4. [SOPs](https://docs.drdroid.io/docs/best-practices-for-writing-sops)

3. **Request & Response Guardrails:**
   1. Isolated AI & backend services – The AI agent can request data but cannot execute actions directly. All execution requests pass through a backend review for correctness & safety.