ALMA Cognitive Engine Guide

The Cognitive Engine provides context tracking, risk assessment, and dynamic persona selection to provide a safer and more intuitive conversational experience.

How it Works

The engine sits in the request loop: User Input -> Cognitive Engine -> LLM Orchestrator -> Execution

1. Frustration Detection & Safety

The system analyzes your input text for keywords that indicate frustration (e.g., "fail", "error", "broken", "wrong").

Frustration Level	System Response
Low (0.0 - 0.3)	Standard operation. Helpful hints enabled.
Medium (0.4 - 0.7)	Terse mode. Removes "fluff", focuses on execution.
High (0.8 - 1.0)	Safety Lock. Destructive commands are blocked. Requires explicit "I confirm" syntax.

2. Adaptive Personas

The AI changes its response style based on the task.

The Architect

Activates when: You ask to create blueprints or design systems. Style: Creative, suggests improvements, explains "why".

"I've designed a Redis Cluster. I suggest adding 3 sentinels for HA. Here is the blueprint..."

The Operator

Activates when: You deploy, rollback, or scale. Style: Military precision. Confirms actions. No chatty text.

"Deploying ID: 592. Status: PENDING. Engine: K8s."

The Medic

Activates when: Things go wrong (errors, logs, troubleshooting). Style: Calm, step-by-step diagnosis.

"I see a 503 error. Let's check the ingress controller first. Logs show..."

3. Context Shifting

ALMA remembers what you are working on.

Example Flow:

1. User: "Deploy web-app-01" -> Context set to web-app-01
2. User: "Scale it to 5 replicas" -> System knows "it" is web-app-01
3. User: "What about the database?" -> Context Shift Detected! Focus moved to database

Configuration

You can tune the sensitivity in .env:

COGNITIVE_FRUSTRATION_THRESHOLD=0.7
COGNITIVE_SAFETY_OVERRIDE=true