RAG evaluation decomposes into three layers: retrieval quality, generation quality, and end-to-end usefulness. Without this decomposition, teams observe “quality dropped” but cannot isolate whether chunking, embedding, retrieval ranking, prompt assembly, or model behavior caused the regression.

The mechanism: each layer has its own metrics, its own failure modes, and its own fix. Retrieval metrics measure whether the right evidence reaches the generator. Generation metrics measure whether the output is faithful to that evidence and actually answers the question. End-to-end metrics measure whether the user’s task got solved. A pipeline can have perfect retrieval but poor generation (model ignores context), or perfect generation but poor retrieval (model faithfully summarizes irrelevant documents).

flowchart LR
    Q[Query + Ground Truth] --> RM[Retrieval Metrics]
    Q --> GM[Generation Metrics]
    Q --> EM[End-to-End Metrics]
    RM --> D1[Did the right chunks arrive]
    GM --> D2[Is the answer faithful and correct]
    EM --> D3[Did the user task get solved]

Example: a support bot returns the correct policy document (retrieval passes) but the model misreads a date constraint and answers with the wrong deadline (generation fails). Without layer separation, the team would chase retrieval improvements that cannot fix a generation problem.

Questions

References

3 items under this folder.