What the documents look like
FBA preparation moves inventory into a platform warehouse ahead of demand. The inbound plan and box labels determine whether goods reach the shelf on time; the inventory report and reimbursement application determine later fees and whether a claim can stand. In store service and internal training, operations staff place inbound plans, box labels, inventory reports and reimbursement applications into a knowledge base so frontline staff can look things up and new hires can study them.
These materials differ in subject, object, stage and validity scope. A retrieval result has to distinguish a draft from a submitted record, a pending review from an approved one, and an application from a completed payout. Materials update daily or weekly. The retrieval layer answers only from materials already present and does not infer states the materials leave unstated.
The inbound plan and box labels sit at the front of the flow, before goods are dispatched. The inventory report and reimbursement application sit later, after receiving and after a discrepancy appears. A single search may return fragments from both ends of that flow, so the reader must check which stage each fragment belongs to.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Retrieval returns reference content as an array, and the array length can be zero, which means the output path still completes when nothing is found.
Semantic retrieval, full-text retrieval, hybrid retrieval and result reranking are optional paths. Hybrid retrieval merges multiple result sets through RRF. Reranking requires a rerank model and must be enabled by configuration. Image retrieval requires image input and the corresponding model capability. The minimum relevance value applies only under semantic retrieval or when result reranking is used; it is a filter threshold, not a final ordering rule, and it does not perform permission or entity isolation. The reference limit is measured in tokens and caps how much content a single search may cite.
Hypothetical paper exercise sample: the knowledge base holds "FBA Inbound Plan — paper sample A", whose text reads "Subject: Xingzhou Daily Goods; Object: batch A-07; Stage: draft; Validity scope: 1 March 2026 to 31 March 2026; Status: not yet submitted." A second item, "Reimbursement Application — paper sample B", reads "Subject: Xingzhou Daily Goods; Object: batch A-07; Stage: submitted, pending review; Claim amount: not stated." The draft sentence to be judged is: "The reimbursement application for Xingzhou Daily Goods batch A-07 has been paid out." This sentence disagrees with sample B's "submitted, pending review", and the sample states no payout fact, so the sentence is unsupported. If the draft sentence instead reads "The inbound plan for Xingzhou Daily Goods batch A-07 has been submitted", it contradicts sample A's "draft, not yet submitted".
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases sharing the vector model of samples A and B | The documents require the same vector model for joint association; check the identifier first, exclude non-matching bases, and do not run cross-model retrieval |
| Search mode | Use full-text retrieval for sample A to check "draft" and "not yet submitted" | Precise status words suit full-text retrieval; compare subject and status against the semantic retrieval result |
| Minimum relevance | Start at 0.5, used only under semantic retrieval or reranking | The value applies only in those two cases; label relevant and irrelevant samples first, then compare whether samples A and B survive filtering |
| Reference limit | Start at 800 tokens, covering the full text of samples A and B | The unit is tokens; compare whether citations stay complete across short and long materials, then decide on adjustment |
| Question optimization | Enable for multi-turn follow-ups, disable for a single clear question | Used for coreference removal and question expansion; compare the original and optimized questions and check whether the subject was altered |
| Result reranking | Enable when a rerank model is available, and filter by its score | Rerank scores run 0 to 1 and are more precise; compare candidate order before and after reranking and confirm sample B is not read as paid out |
Three common errors
- Treating retrieval recall as authorization to act. Sample B sits at "submitted, pending review". Finding that material shows only that the content exists; it does not show the reimbursement was approved or paid. A pending review is not an approval, and an application is not a completed payout.
- Associating knowledge bases that use different vector models. The documents require the same vector model for joint association. Check the identifier and exclude non-matching bases; do not invent missed recalls or redundant results from a cross-model association.
- Treating minimum relevance as a permission or entity filter. The value is a similarity threshold and does not exclude unrelated business lines or unauthorized records. Retrieved results still require manual checking of subject, object, stage and validity scope.
How to verify the setup
- Input sample A's text and the draft sentence "The inbound plan for Xingzhou Daily Goods batch A-07 has been submitted", then compare whether the draft sentence contradicts the original's "draft, not yet submitted". A contradiction means the draft sentence is unsupported, which is the expected observation.
- Input sample B's text and the draft sentence "The reimbursement application for Xingzhou Daily Goods batch A-07 has been paid out", then compare whether "submitted, pending review" and "paid out" are mutually exclusive values of the same field. The sample states no payout fact, so the result should be unsupported or pending verification, not a denial or a conflict.
- Input an empty reference array [] together with the downstream design, then compare whether the downstream path continues on an empty result. The documents state the reference length can be zero and the path still completes; if the downstream requires non-empty content to continue, record that as a design difference. All three checks are paper comparisons and do not claim any test, deployment or approval took place.