Building AI.
Scaling Systems.
Leading Engineering.

Retrieval, long context or fine-tuning?

A way to ask the right questions, not a benchmark. Your data, volume and constraints decide.

The knowledge changes often, or is private to you.

Retrieve it at request time. Keep the index easy to update, filter by who is asking, and show the sources.

The whole body of knowledge is small, stable and fits comfortably in the context window.

Try putting it in the prompt first. Add retrieval only when it stops fitting or the cost becomes a problem.

You need a consistent format, tone or behaviour that prompting does not deliver.

Consider fine-tuning, but only after an evaluation set shows that prompting and retrieval fall short.

You cannot yet say how you would know it works.

Build the evaluation set first. Every other choice here depends on it.

Go deeper

• Request pipeline in depth: Designing the AI Request Pipeline: 8 Layers Between User Input and Your LLM
• Memory for long-running agents: Agent Memory Patterns: How to Give Your LLM Application a Brain That Persists
• Shipping model changes safely: Feature Flags for AI: The Deployment Pattern That Saved Us From 3 Production Disasters