RAG answers are poor. What should you check first?
Why A is the answer
Whether retrieval returns the right chunks
If retrieval fails, the model never sees the right content.
In detail
- If retrieval is wrong, nothing later can fix it.
- A model cannot use text it never received.
Use case
An illustrative scenario, not a real incident.
RAG answers are poor. The team checks the retrieved chunks first, and finds that the right text never arrives.
See it in different ways
flowchart TD
A["Poor answers"] --> B{"Right chunks retrieved?"}
B -->|No| C["Tune chunking, embeddings and search"]
B -->|Yes| D["Tune the prompt or model"]stateDiagram-v2 [*] --> CheckRetrieval CheckRetrieval --> TuneChunks: wrong chunks CheckRetrieval --> TunePrompt: right chunks, poor answers TunePrompt --> FineTune: style still not met TuneChunks --> [*] FineTune --> [*]
flowchart TD
Q{"Are the right chunks retrieved?"}
Q -->|"No"| A0["Fix retrieval"]
Q -->|"Yes"| A1["Tune the prompt"]Key point: Check retrieval first.
Why the other options are wrong
- B. Whether a larger model is available this month
- A larger model cannot read missing chunks.
- C. Whether fine-tuning can be started today
- Fine-tuning is not the first step.
- D. Whether the temperature could be raised
- Temperature adds no facts.