What the documents look like
Third-party overseas warehouses can serve several sales channels at once, and the seller allocates stock across those channels. Replenishment planning, inbound appointment requests, inventory age tables and final-mile delivery notes form the working document set. Records arrive from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems and carrier statements, with weekly or daily refresh. Contract and transaction material comparison needs retrieval results that map to a specific party, object, period and field. The retrieval step returns candidate content only; it does not decide authorization.
For this page the comparison question is narrow: does a proposed sentence match what the source materials actually record about party, object, period and status. A retrieval hit on a replenishment plan or an inbound appointment shows that content was recalled. It does not show approval, completion or permission.
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. During checking, confirm the vector model identifier of each base, exclude those that do not match, and record what was observed. Do not run a cross-model search to compare missed recalls.
Search modes include semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process inside hybrid retrieval, not a switch that is turned on. The minimum relevance value applies only under semantic retrieval or when result reranking is used. That value is a filtering threshold, not a permission or entity isolation mechanism. The citation limit is measured in tokens and controls how much content a single search cites.
Question optimization performs coreference resolution and question expansion, and it adds a model call and extra latency. Image retrieval requires image input and matching model capability. Reranking requires a rerank model and must be enabled in the settings. Optional paths in the overall retrieval flow do not all execute on every request.
Hypothetical paper exercise sample. A replenishment plan material reads: "Party: a cross-border seller in East China; Object: SKU-A replenishment plan for the Bremen warehouse in Germany; Period: week 10 of 2025; Quantity: 600 units; Status is pending review." An inbound appointment material reads: "Party: a cross-border seller in East China; Object: SKU-A, Bremen warehouse in Germany; Appointment inbound date: 5 March 2025; Status is submitted." The draft sentence to be assessed is: "The SKU-A replenishment plan has been approved, and 600 units will enter the warehouse on 5 March 2025." When this draft sentence is submitted as the retrieval input, the plan material records status pending review and the appointment material records status submitted. Neither material states approval or entry completion, so the approval and inbound statements in the draft sentence are not supported.
Configuration and implementation checks
| Configuration item | Purpose on this page | Setting or checking method |
|---|---|---|
| Linked knowledge bases | Recall replenishment plan, inbound appointment, inventory age and delivery note materials | Confirm each base uses the same vector model; exclude mismatched identifiers and record the observation |
| Search mode | Match party, object, period and status fields | Compare semantic, full-text and hybrid retrieval on the test question; do not assume one mode fits all materials |
| Citation limit | Control cited content per search | Unit is tokens; test content completeness with materials of different lengths before adjusting |
| Minimum relevance | Filter low-relevance candidates | Applies only under semantic retrieval or result reranking; label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Complete references and expand the question | Compare the original question with the optimized question and check whether the party was altered |
| Result reranking | Reorder candidate text a second time | Requires a rerank model and must be enabled in the settings; the score is used for filtering |
A hypothetical starting value for the citation limit in this exercise is 2,000 tokens, chosen because the replenishment plan and inbound appointment excerpts together fit well below that size while remaining readable end to end; it is compared against a smaller value to see whether the appointment date field is cut off. This is an exercise starting point, not an official default.
Three common errors
- Treating a retrieval hit as authorization confirmation. In the sample, the replenishment plan status is pending review. Recalling that material shows the content was returned; pending review does not demonstrate a lack of operating permission and does not constitute approval. Check: submit the draft sentence, compare the status field in the materials, and observe whether any approval or entry statement appears.
- Linking knowledge bases that use different vector models. Check: inspect the vector model identifier of each base, exclude those that do not match, and record the observation. Do not perform a cross-model search to compare missed recalls.
- Merging quotation validity, signing, cooperation reached and contract effectiveness into one fact. In the sample, the inbound appointment status is submitted. Submitted does not equal warehouse entry completed. Check: submit the appointment material text, compare the status field, and observe whether any completion or effectiveness statement appears.
How to verify the setup
- Submit the hypothetical draft sentence "The SKU-A replenishment plan has been approved, and 600 units will enter the warehouse on 5 March 2025" and compare it against the plan material status pending review and the appointment material status submitted. Condition satisfied when retrieval returns both materials and the comparison conclusion states that the approval and inbound statements are not supported, with the evidence gap preserved.
- Submit the paper input [] and check how the downstream design handles an empty reference. Condition satisfied when the downstream chain still runs with a reference length of 0, without claiming that a fabricated rule search necessarily returns [].
- Submit a short question containing only the party and the object, and compare the retrieval terms before and after question optimization is enabled. Condition satisfied when the optimized question keeps the original party and object, does not alter the party, and the retrieval results can be checked against the corresponding material fields.