Copyright (c) 2026 MindMesh Academy. All rights reserved. This content is proprietary and may not be reproduced or distributed without permission.

7.2. Quick Reference

The Master Decision Filter

Run this on every scenario before reading the options closely.

Deployment Type Quick-Reference

The Lever Selection Table (highest-value distinction)

NeedLeverNever use
Steer one response (tone, format, this turn)Prompting
Supply private/fresh factsGrounding (RAG)Fine-tuning (won't add facts)
Consistent style/format across callsFine-tuningGrounding (it's not facts)
Take an action / call a systemFunction/connected toolCode Interpreter (sandbox only)
Analyze data, run math, make chartsCode InterpreterFunction tool for built-in analysis

Control-to-Threat Quick Map

Threat / needControl
Runaway token spend (prevent)TPM ceiling on deployment
Spend awareness (notify)Budget alert
Jailbreak / prompt injectionPrompt Shields
Harmful content (hate/sexual/violence/self-harm)Harm-category filter + severity
Model invents facts not in sourceGroundedness detection
Sensitive data before an LLMLanguage PII redaction (preprocess)
Authenticate without secretsManaged identity (DefaultAzureCredential)
Over-permissioned runtime identityLeast-privilege RBAC (OpenAI User)
Control/audit encryption keysCustomer-managed keys (CMK/BYOK)

Search Strategy Quick-Reference

Query traitUse
Exact codes/IDs + conceptsHybrid (keyword + vector)
Concept/paraphrase onlyVector
Need top-result precision/nuanceAdd semantic ranking (cross-encoder)
Keep index fresh, no custom codeIndexer + integrated vectorization
Alvin Varughese
Written byAlvin Varughese
Founder18 professional certifications