What the documents look like
FBA preparation moves inventory ahead of demand into platform warehouses, so inbound discrepancies and inventory age are the main risks. Order fulfilment and after-sales answering therefore revolve around four material types: inbound plans, box labels, inventory reports and compensation claims. Operations staff, overseas warehouse logistics specialists and purchasing each maintain part of this material, with updates occurring daily or weekly.
The retrieval step has to cope with version and status differences across these materials. An inbound plan for a shipment may already be confirmed while the corresponding box label file is still pending review. An inventory report may show a received quantity that differs from the planned quantity. A compensation claim may still be under platform review. Retrieval results present document text only. They do not determine whether an authorization state has been reached, and a similarity score does not act as a permission or entity filter.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model. The smallest search unit is the library; a collection serves only as a classification layer and does not affect search results. A single retrieval combines semantic search, full-text search, hybrid search, question optimization and reranking according to the input content and search parameters, then applies similarity filtering and a citation limit before returning a citation content array whose length may be zero. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Hybrid search merges multiple result paths through RRF, and rerank scores can be used for filtering. Image retrieval requires corresponding image input and model capability, and reranking requires an available model enabled in settings; these paths do not run on every request.
Hypothetical drill sample: the source text for one FBA inbound plan reads "Plan number H-01, shipped quantity 600 units, destination warehouse North America, status confirmed; box label file B-01 status pending review; inventory report R-01 records 540 units received, difference of 60 units awaiting verification; compensation claim C-01 status under platform review." The draft sentence to be judged is "The 60-unit difference for batch H-01 has received platform compensation." This sentence concerns subject H-01, object the 60-unit difference, and status compensation received, which conflicts with the source wording "awaiting verification" and "under review." It is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge libraries | Associate only libraries carrying the four material types above that share one vector model | Check each library's vector model identifier first, exclude those that do not match, and do not associate across models |
| Search mode | For identifier-type questions, compare full-text search with hybrid search | Use identifiers such as H-01 and B-01 as test questions and observe whether the returned fragment contains the matching identifier |
| Citation limit | Start at a token amount that can hold the complete fields of one inbound plan | Test with actual materials of different lengths for content completeness, then compare against neighbouring values |
| Minimum relevance | Start at 0.5, applied only to semantic search or reranking | Label relevant and irrelevant samples first, then compare filtering behaviour at different thresholds |
| Question optimization | Enable for multi-turn follow-ups, compare with it disabled for questions containing explicit identifiers | Check whether the subject of the optimized question has been altered incorrectly |
| Empty-result handling | Design the downstream branch around a citation content array of length zero | Use a paper input of an empty array to check the downstream branch; do not claim that any rule necessarily returns an empty array |
Three common errors
- Treating a box label pending review as approved and scheduling inbound on that basis. Input: the draft sentence "B-01 box label approved, inbound can be scheduled." Comparison object: the source wording "status pending review." The condition is met only when the source text contains a record of review approval; otherwise the item is recorded as unsupported.
- Writing a compensation claim under platform review as compensation received. Input: "C-01 compensation received." Comparison object: the source wording "under review." The condition is met only when the source text contains a record of compensation completion or payment arrival.
- Using a similarity score to exclude unrelated business records. Input: "filter out other stores' data by minimum relevance." Comparison object: the documented nature of the filtering threshold. The condition is met only when an explicit permission field and filtering implementation exist; the retrieval score itself does not perform that function.
How to verify the setup
- Compare the draft sentence "The 60-unit difference for batch H-01 has received platform compensation" field by field against the source wording "difference of 60 units awaiting verification" and "C-01 status under platform review." Input: the draft sentence and the source text. Compare the subject, object and status. The condition is met only when the source text shows a matching status of compensation received; this record is unsupported.
- Use identifier H-01 as a test question and observe the returned citation content under full-text search and under hybrid search separately. Input: the identifier question. Compare whether the returned fragment contains H-01 and the quantity 600. A fragment containing both the identifier and the quantity meets the condition; a miss is recorded as pending.
- Use a paper input of an empty array to check the downstream branch. Input: an empty citation content array. Compare whether the downstream continues to execute along the path whose length may be zero. Normal downstream handling without error meets the condition; when actual behaviour has not been observed, record it as unknown.