AI shopping assistants increasingly employ agent-based retrieval, combining lexical search, structured filtering, and LLM-mediated selection. However, effective retrieval often requires knowledge beyond textual matching. This exploratory single-case study of an Indonesian grocery e-commerce assistant triangulates 279 observations, 52 failure traces, 74 practitioner-reported defects, a schema audit, and four practitioner interviews. The analysis identifies six retrieval-critical knowledge dimensions and reveals that externalization fails at two distinct layers. At the product data-model layer, essential fields (allergens, dietary constraints, age suitability) were absent and persisted despite architectural changes. At the retrieval-schema layer, existing knowledge failed to reach candidates: structured filters appeared in only 29% of calls, and correctly invoked filters often returned empty sets from non-empty pools. Addressing these layer-specific failures, the study proposes a structured knowledge framework based on the knowledge management process cycle, positioning GraphRAG as a future direction.