What it is
LlamaIndex started as the data layer: ingest documents, chunk them, embed them, index them, retrieve against them. Agents and workflows came later and sit on top of that retrieval core.
What goes wrong
RAG failures are almost always retrieval failures wearing a generation costume. The model says something wrong and the instinct is to blame the model, when the retriever handed it three irrelevant chunks and it did about as well as anyone could with them.
What the adapter does
Attaches to the dispatcher event system and links each generated span back to the specific retrieval that fed it. When output is wrong you can look at what was actually in context rather than reasoning about what should have been.
adapter = LlamaIndexAdapter("app"); adapter.attach()
Caveat
Retrieval quality is upstream of everything here. Causal tracing tells you the retriever is the problem. It does not tell you your chunking strategy is the reason.