Method
Retrieval, fine-tuning, RL: a field guide to adapting open models
Knowledge becomes retrieval, skill becomes fine-tuning, outcome becomes optimization. A practical map of how we turn the best open models into a profession for a specific business.
We don't build the brain. The frontier open models are extraordinary general reasoners — our job is to turn one into a specialist that knows your business and is measured against your number. There are three levers, and most real systems use all three.
1. Knowledge → Retrieval (RAG)
The model doesn't know your policies, your part catalog, or last quarter's incident reports. Retrieval gives it access at answer time.
- Ground responses in your documents, not the open web.
- Update knowledge by updating the source, not by retraining.
- Cite, so a human can verify the chain.
Retrieval is the fastest lever and the one most teams under-invest in.
2. Skill → Fine-tuning (LoRA)
Some things can't be retrieved — they have to be learned. The shape of your outputs, your formatting standards, the judgment calls your best operators make.
Low-rank adaptation lets us teach that behavior efficiently, without touching the base weights. We fine-tune for style and skill, not for facts — facts belong in retrieval.
3. Outcome → Optimization (RL)
The last lever is the hardest and the most valuable: optimize the model against the metric that actually matters.
Reinforcement learning drills the system toward a defined outcome — fewer escalations, higher yield, lower cost-per-unit — until the behavior is the result you wanted.
Putting it together
A production system is rarely one lever. It's retrieval for knowledge, fine-tuning for skill, and optimization for outcome — composed, evaluated, and run in our own operations before it's yours.