Postgres RAG chunking boundaries shape retrieval quality in knowledge-base bots

An internal knowledge-base bot showed degraded answer quality even though its services stayed up and queries succeeded. The trace revealed that the cited sources pointed to the correct handbook but the wrong passage, meaning the retrieval unit itself was flawed. Chunks that mix topics match weakly, while chunks that split a procedure across boundaries may omit the answer entirely. The guidance recommends structure-aware boundaries such as headings and complete procedures, paired with an evaluation set that checks whether the top results contain the supporting passage. Alerting should focus on sustained losses of answer-bearing chunks in retrieval results rather than on fluency of model output.