Error propagation analysis tracks how a failure at one stage affects later processing and the final response. In large, single prompts, one misunderstood requirement can make the entire output unreliable. Composable, multistep systems isolate failures and support targeted recovery. section 12.2.2
Analyze the reasoning chain stage by stage:
- Check retrieval: determine whether the system found relevant information for the query. section 12.4.4
- Check context assembly: determine whether retrieved information was combined logically and appropriately. section 12.4.4
- Check generation: determine whether the final response reflects failures inherited from earlier stages. section 12.4.4
- Trace token allocation and branching points to identify inefficient context use or decisions that redirected the reasoning chain. section 12.4.4
Tracing a RAG failure
A question about vacation policy may fail because retrieval selects the wrong document chunks, context assembly omits a relevant section, or generation misinterprets the assembled evidence. Langfuse span tracking exposes the flow through retrieval, augmentation, and generation, including timing and token usage, so the failing stage can be investigated directly. section 12.2.2 section 12.4.4
