RAG Stage Blueprint from Loading to Evaluation
Lay out all five RAG stages with the concrete component choices and success measure for each.
A five-stage blueprint (loading, indexing, storing, querying, evaluation) with component choices, rationale and a measurable check per stage.
Role
You are producing an implementation blueprint for a retrieval-augmented application.
Task
Blueprint a RAG system for {{use_case}}. Stage Loading: choose connectors from {{connector_list}} and define how source documents become nodes using {{node_strategy}}, including the metadata that links nodes back to their document. Stage Indexing: select {{index_type}} and state what is embedded versus stored as metadata. Stage Storing: use {{storage_target}} and describe what is persisted so re-indexing is unnecessary. Stage Querying: implement {{query_strategy}}, stating whether sub-queries or multi-step retrieval are used and how postprocessing and response synthesis are configured. Stage Evaluation: define {{evaluation_metrics}} with a target and the measurement procedure. Give one measurable check per stage.
Context
Treating RAG as five explicit stages localises regressions: a drop in answer quality is traceable to loading gaps, index configuration, retrieval strategy or synthesis rather than being diagnosed as one blob.
Inputs
- {{use_case}}
- {{connector_list}}
- {{node_strategy}}
- {{index_type}}
- {{storage_target}}
- {{query_strategy}}
- {{evaluation_metrics}}
Constraints
- Every node must retain a link to its parent document
- State what is persisted versus recomputed
- Query strategy must say when it escalates to multi-step retrieval
- Every stage needs at least one measurable check
Output Format
Markdown with one section per stage plus a stage-check summary table.
Quality Criteria
- Component choices justified against the use case
- Node-to-document traceability preserved
- Evaluation metrics quantitative
- Checks would localise a quality regression