Architecture choice
Do not add retrieval when instructions and selected files are enough
Prompts define how work is performed. Files provide selected knowledge. RAG retrieves passages from larger controlled collections. They solve different problems.
- Reviewed
Three options
Match complexity to the actual knowledge task
- Prompt-first
- Use for a few stable tasks where method and output format are more important than a document collection.
- Repository or file-first
- Use for reusable methods, templates and selected knowledge that can still be deliberately supplied.
- RAG-first
- Use when document volume, citations, updates, access or repeatable natural-language search require controlled retrieval.
Decision questions
RAG is justified by control requirements
- How many sources exist and how often do they change?
- Must answers cite exact passages?
- Do different users have different access?
- Can removed documents be proven absent?
- Can representative retrieval questions be evaluated?