What the documents look like
Third-party overseas warehouses can serve several sales channels at once, and the seller allocates stock across those channels. The working material for replenishment therefore spans more than one system. Replenishment plans, inbound appointment requests, inventory age reports and last-mile delivery orders come from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems and carrier bills. These records update weekly or daily, so a retrieval answer can be current for one document and stale for another.
For product data and selection questions, the retrieval step has a narrow job. It returns quoted content that carries a subject, an object, a status and a validity range, and that content is then checked by a person. Retrieval does not decide whether an operation is authorized. A retrieved record that says a replenishment plan is pending review remains a pending record, and no search result converts it into an approval.
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. The practical step is to read the knowledge base identifier and the vector model identifier, compare them, and exclude any knowledge base whose model does not match. Cross-model retrieval is not run as a comparison method.
Search mode is a comparison decision rather than a fixed setting. Semantic retrieval, full-text retrieval and hybrid retrieval are tested against the same question, and the recalled content is compared across modes. No single mode is assumed to suit all replenishment material. Full-text retrieval suits identifiers, model numbers and proper nouns; semantic retrieval suits natural-language questions and cross-language matching.
The reference limit is expressed in tokens, not in a top-k item count. Because chunk lengths differ widely in a mixed knowledge base, a token limit is used to control how much content is quoted. Long and short source documents are compared to see whether the quoted content is truncated. The minimum relevance value is a number between 0 and 1 and takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final ordering rule, and it does not act as a category, entity or permission filter. A similarity score does not exclude unrelated businesses or unauthorized records.
Question optimization performs coreference resolution and question expansion before retrieval. It adds a model call and therefore adds latency. The original question and the optimized question are compared to check whether the subject has been altered. Reranking requires a rerank model and is enabled according to the configuration; it produces a score between 0 and 1. RRF is the fusion step that combines multiple retrieval paths, and it is not an on/off switch. Image retrieval requires image input and the relevant model capability, and reranking requires a model; the optional paths in the overall flow are not all executed on every request.
Assumed paper exercise sample: a replenishment management record reads "Replenishment plan PL-2025-031, object German warehouse SKU-A, status pending review, valid until 30 April 2025. The inbound appointment must be submitted 48 hours before arrival at the warehouse." A draft sentence reads "This replenishment plan has been approved and the inbound appointment can be arranged directly." After retrieving the record, the draft sentence disagrees with the source status field. The source says pending review; the draft says approved. The draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Assumed starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the replenishment knowledge base that uses the same vector model | Read the knowledge base identifier and vector model identifier, compare them, and exclude bases whose model differs |
| Search mode | Run one round with semantic retrieval and one round with full-text retrieval | Use the same test question and compare the recalled content across the two rounds |
| Reference limit | Assumed starting value 800 tokens | Compare one long and one short source document to see whether quoted content is truncated |
| Minimum relevance | Assumed starting value 0.5 | Label relevant and irrelevant samples first, then compare the filtered results |
| Question optimization | Run one round with it off and one round with it on | Compare the original question and the optimized question to see whether the subject was altered |
| Result reranking | Run one round with it enabled when a rerank model is available | Compare the candidate ordering before and after reranking |
The reference limit and minimum relevance values above are starting points for this paper exercise, chosen so that a comparison can be made. They are not documented defaults. The associated knowledge base requirement comes from the search node description; the token unit and the minimum relevance condition come from the search parameter documentation.
Three common errors
- Treating retrieval as authorization. A retrieved replenishment plan with status pending review is not an approved plan. Input the source status field, compare it with the status word in the draft sentence, and observe whether the two agree. In the sample, pending review and approved disagree, so the draft sentence is unsupported.
- Associating knowledge bases that use different vector models. Input the knowledge base identifier and the vector model identifier, compare them, and observe whether they match. When they differ, exclude that knowledge base and do not run a cross-model retrieval comparison.
- Treating minimum relevance as a permission or entity filter. Input an unrelated business query and observe whether quoted content can still be returned. The relevance score does not perform the function of excluding unauthorized records.
How to verify the setup
- Input the assumed source text "Replenishment plan PL-2025-031, status pending review" and compare it with the draft sentence "has been approved." Observe whether the status fields agree. When they disagree, the draft sentence is not supported. A pending record is not an approval, and no further authorization clause should be invented to make the answer definite.
- Input the same test question once with semantic retrieval and once with full-text retrieval. Compare whether the recalled content covers the source subject, object, status and validity period. When one mode omits a key field, record that gap. Do not assume that either mode returns every relevant item or that the first result is correct.
- Input an empty-reference case as the paper input [] and check the downstream design against it. Observe whether the output path still completes when no content is retrieved. Do not claim that a fictional rule search necessarily returns [].
Across all three checks, the comparison must use the same subject, the same period and the same field. Only mutually exclusive values in the same field count as a contradiction. A status that is not stated, or evidence that covers only a related stage, is unsupported or awaiting verification. Retrieval is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where the source material needed for a check is not present in the accessible retrieval scope or in a stated manual input, record the gap and leave it open.