What the documents look like
Third-party overseas warehouses serve several sales channels at once, and the seller allocates stock across those channels. The material behind replenishment work arrives from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems and carrier statements. The documents handled day to day are replenishment plans, inbound appointment requests, inventory age reports and last-mile delivery notes. In store service and internal training, these documents enter a knowledge base so frontline staff can retrieve replenishment rules, appointment procedures and the basis for inventory age charges. Retrieval results feed judgements about stockouts, slow-moving storage rent and last-mile delay, so recalled content has to stay consistent with the source document wording, and the reference cap and minimum relevance settings need checking against actual document length and question type. The brief gives weekly and daily update frequencies; those frequencies do not change how retrieval behaves on this page, so they are not carried further.
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. At retrieval time, vectors are recalled first, then the vector ID is used to find the original data in the MongoDB dataset.datas table; when several recall paths hit the same original data group, the results merge and the vector score takes the highest value. Hybrid retrieval combines semantic and full-text results through RRF, and RRF is a fusion process rather than a switch. Minimum relevance takes effect only in semantic retrieval or when result reranking is used, and it is a filter threshold rather than a ranking rule. The reference cap is measured in tokens; because long and short material is mixed in the same base, top k is unstable, which is why token control is used. Question optimization performs coreference resolution and question expansion and adds one model call. Retrieval itself does not carry permission or entity isolation, and a similarity score does not distinguish business ownership.
Assumed drill sample: an overseas warehouse operations staff member submits one request for judgement during internal training. The drafted sentence is "This batch replenishment plan has been approved by the store manager, so the inbound appointment can be started directly." The source text supplied for comparison reads: "Replenishment plan number BH-2024-0312, submission date 12 March 2024, current status is awaiting store manager review. The inbound appointment must be initiated after approval." The request carries no approval record and no signed page. This sample is used to check whether the drafted sentence and the source text agree on status.
Configuration and implementation checks
| Configuration item | Assumed trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Select the inventory age base and the inbound appointment base | First compare the vector model identifiers of the two bases; if they differ, exclude them and do not run a cross-model search |
| Search mode | Test semantic retrieval and full-text retrieval separately | Compare full-text hits on number-type questions and semantic hits on natural-language questions |
| Reference cap | Assumed starting point of 2000 tokens | Test whether inventory age passages of different lengths are truncated, comparing against the full passage length |
| Minimum relevance | Assumed starting point of 0.5 | Label related and unrelated samples first, then observe whether the filtered references keep the subject information |
| Question optimization | Enabled for multi-turn follow-ups, disabled for single clear questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when a rerank model is available | Compare the rerank score with the vector score and observe changes in candidate ordering |
These values are assumed trial starting points for this page and are not official defaults.
Three common errors
- Treating an awaiting-review status as approved. The sample source text states "current status is awaiting store manager review" while the drafted sentence states "has been approved by the store manager"; the two values for the same field are mutually exclusive, which forms a contradiction. To check, input the drafted sentence and the source text, compare the status field, and if the source says awaiting review while the draft says approved, the draft does not hold.
- Treating a retrieval hit as authorized execution. Returned reference content only shows that text was recalled; it does not show that approval passed or that an appointment was initiated. To check, input the retrieval request and the downstream appointment request, compare whether a separate approval record exists, and if only retrieval output exists without an approval record, the appointment request remains unsupported.
- Linking knowledge bases that use different vector models. The documentation states that only knowledge bases with the same vector model can be linked. To check, input the model identifiers of the two bases, compare whether they match, and if they do not, exclude them, do not run a joint search, and do not infer what a joint run would produce.
How to verify the setup
- For the drafted sentence in the assumed sample, input the source text "current status is awaiting store manager review" and the draft "has been approved by the store manager", compare the values of the status field, and if the two are mutually exclusive, the draft does not hold; record it as pending verification and do not supply an approval record.
- For the empty-result path, input an explicit paper input of
[]and check whether the downstream design lets the chain run through with a reference length of 0; do not claim that a fictional rule search will necessarily return[]. - For cross-model linking, input the vector model identifiers of the two bases and compare whether the identifiers are the same; if they differ, exclude the non-matching base and record the actual observation, and do not run a cross-model search to compare missed recalls.
This page covers only overseas stock replenishment management, store service and internal training, and knowledge base retrieval and recall. When operations are handed over, which materials, actions and requests sit at which authorization state cannot be treated as authorized execution on the strength of a retrieval hit or a form entry.