Fine-Tuning vs RAG Consulting

Fine-tuning vs RAG consulting is a decision and build engagement: whether your task needs retrieval, a weight update, or both. We score prompt-only, RAG, and PEFT variants on your golden set, then implement the winner in your cloud. You own the IP. Zero-retention, no shared training, SOC 2-aligned. Four weeks is standard; financial-services review typically takes 8–12 weeks.

When the job is answering from documents that change, must be cited, or are permissioned per user. RAG retrieves at query time, so a policy update is an index refresh, not a retrain. Fine-tuning those facts into weights makes them stale and cannot enforce query-time ACLs. If someone says “train on the SharePoint dump,” the default answer is RAG.

When prompting and RAG still miss a skill: house format, tool-call reliability, language, or tone that few-shot examples cannot hold. You need a task dataset, a held-out eval split, and a base model you are licensed to adapt. Fine-tuning is not how you store facts. If we cannot name the behavior the weights must learn, we will not start a train job.

When the model must behave in a specific way and also ground answers in current, permissioned documents. Typical pattern: PEFT for schema and tone, RAG for facts and citations. Both paths still need golden sets, rubric judges, and CI gates. Combining them without evals just stacks two ways to fail.

Usually no. Dumping the wiki into continued pretraining teaches the model to parrot pages, not to complete your job, and it still cannot cite or respect ACLs. Convert work into input/output pairs that match production prompts and tools if you fine-tune. Keep facts in retrieval. We will say this in the decision memo when a dump is the proposed dataset.

You do. Training sets, LoRA adapters, vector or hybrid indexes, prompts, evals, and traces are client IP in your cloud. We do not retain them and we do not use them for shared training. Base-model licenses still apply to the weights you start from. The process is SOC 2-aligned.

A golden set of real cases, including empty-retrieval and must-refuse items, scored with deterministic checks and calibrated rubric judges. We report retrieval recall, faithfulness, format validity, and task success as separate numbers. CI gates the winning path so a later prompt tweak or adapter bump cannot silently undo the decision.

Four weeks is the standard path to a scored decision and a first production implementation of the winner. A heavy PEFT dataset or a regulated RAG corpus can extend build. Financial-services programs with second-line review typically take 8–12 weeks. GPU wait time sits outside the consulting clock if you have no training capacity yet.