How do you troubleshoot poor retrieval relevance in a LlamaIndex application?

When a RAG application returns answers that miss the point or cite the wrong context, the fix usually starts by actually inspecting the source_nodes on the response object, rather than guessing, since that tells you whether the problem is retrieval or synthesis.

1. Check what was actually retrieved. If the correct Node never shows up in source_nodes, the problem is retrieval, not the LLM.
2. Re-examine chunk size. Overly large chunks dilute embeddings; overly small ones lose context. Try adjusting chunk_size/chunk_overlap and re-testing.
3. Check for vocabulary mismatch. If queries and documents use very different phrasing, a query transform like HyDE can help.
4. Verify metadata filters aren't over-restrictive, silently excluding the correct Node.
5. Adjust similarity_top_k up if relevant Nodes are close but just outside the current cutoff, or add a reranker if the right Node is in a larger candidate set but ranked too low.
6. Confirm ingestion actually completed correctly, since a silently failed or truncated load will produce an index missing content entirely, which no amount of query tuning will fix.
7. Reconsider the index type. A question needing information from across an entire document may be poorly served by a VectorStoreIndex's top-k retrieval and better handled with a SummaryIndex or tree_summarize.