What the documents look like
FBM fulfillment is run by the seller. The seller arranges delivery and sets the return address, and the platform warehouse does not inspect the goods. For store service and internal training, the knowledge base therefore holds four document types: order picking lists, shipment confirmation forms, return address settings, and logistics tracking sheets. These materials update per order or daily, so a single order number can appear at four different stages — picking, shipment, return, and tracking — and the records differ at each stage.
Staff in customer service and logistics handovers often need to confirm which state a given document is currently in. A retrieval result is quoted content only. It does not grant authorization and does not execute any business action. Whether a shipment confirmation has actually been signed, whether a return address change has been approved, and whether a tracking record has been accepted by a carrier are separate facts that the quoted text may or may not establish.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. Result reranking requires a rerank model and must be enabled by configuration. The citation cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. The quoted content is an array whose length can be zero, and the downstream chain still runs when nothing is retrieved.
Retrieval and similarity scores do not carry permission or entity isolation duties. Business authorization must be implemented downstream. A similarity setting does not exclude unrelated businesses or unauthorized records.
Hypothetical paper exercise. An FBM seller's internal training material states: "Order A-1024 has been picked; the picking list status is printed. The shipment confirmation form has not been signed. The return address setting is the default overseas warehouse address. The logistics tracking sheet shows the carrier has collected the parcel." A draft sentence under review reads: "Order A-1024 has completed shipment confirmation, and the return address has been changed at the customer's request." The review scope is limited to A-1024 materials in this seller's internal knowledge base. No real customer or external system is involved.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the store service base that uses the same vector model | Check each base's vector model identifier first, exclude mismatches, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Compare recalled content for the same question against semantic retrieval and full-text retrieval separately |
| Citation cap | 800 tokens | A starting point for this exercise; compare against 400 and 1200 for content completeness |
| Minimum relevance | 0.35 | A starting point for this exercise; label relevant and irrelevant samples first, then adjust |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled | A rerank model must be available; compare ordering before and after reranking |
These values are starting points for the paper exercise, not official defaults. The comparison method for each row is a manual review step, not an automated control.
Three common errors
- Treating a retrieval result as an authorization basis. Input: the draft sentence "Order A-1024 has completed shipment confirmation." The knowledge base source shows only that the shipment confirmation form has not been signed. Comparing the same field across the two texts, the source does not state that signing was completed. The draft sentence is unsupported, and no shipment confirmation action may be executed on that basis.
- Linking knowledge bases across different vector models. Input: two bases whose vector model identifiers differ. After checking the identifiers, exclude the mismatched one and record what is observed. Do not fabricate an outcome for a joint link, and do not describe a cross-model search as producing missed or redundant results.
- Treating a pending review state as approved. Input: a return address change request that is pending review. The source does not show approval. The request does not constitute an approved or executed change. A pending request is not approval.
How to verify the setup
- Input the hypothetical source text and the draft sentence. Compare "the shipment confirmation form has not been signed" with "has completed shipment confirmation." The source does not state that signing was completed, so the draft sentence is unsupported, and the retrieval result must not be used as an authorization basis for execution.
- Input an empty result array []. Check whether the downstream design accepts a quoted-content chain of length zero. Do not claim that a fictional rule search necessarily returns [].
- Input the same question under semantic retrieval, full-text retrieval, and hybrid retrieval. Compare the three recall paths and the ordering after RRF merging, and observe score changes before and after reranking is enabled. Do not guarantee a first-position hit or complete recall.
Each check states what is compared and what remains unknown when the source is silent. Missing evidence stays missing.