Key takeaways up front
- Learn retrieval first — it covers most business problems
- Fine-tuning fixes format and tone, not knowledge
- Chunking and evaluation cause most production failures
The question is asked as if it were a rivalry. In practice retrieval handles knowledge and fine-tuning handles behaviour, and most teams need the first long before the second.
1The decision tree
Answer these in order and the choice usually makes itself.
- Does the answer depend on data that changes? → retrieval
- Do you need a consistent output format or tone? → fine-tuning
- Is your corpus under a few thousand documents? → retrieval, no question
- Do you have labelled examples in the thousands? → fine-tuning is viable
2Retrieval mistakes that fail review
These are the issues that show up in every production readiness review.
- Fixed-size chunking that splits tables and clauses
- No metadata filters, so old documents outrank current ones
- No evaluation set, so quality claims are anecdotal
- No citation surfacing, so users cannot verify answers
3What to build to learn it properly
Take a messy real corpus — policy documents, invoices, support tickets — and build retrieval with an evaluation set of 50 question-answer pairs. That single project teaches more than any course module.
What to do next
- Retrieval first, fine-tuning only when behaviour is the problem.
- Chunk along document structure, never by character count alone.
- No evaluation set means no evidence — build it before the demo.
#RAG#AI agents#Prompt engineering