What the documents look like
Parcel label and handoff work sits at the point where a parcel leaves the warehouse. The carrier's scan action defines the core responsibility at this stage. Compared with in-warehouse packing, the material set here centers on electronic waybills, handoff lists, pickup scan records, and exception parcel registrations. Sources include the courier system, WMS shipment records, GB/T courier service standards, and carrier billing. Updates arrive per batch and per day, so retrieval has to cope with the same batch of documents carrying different status values at different times.
This page covers one task only: knowledge base retrieval and recall under operations permission and handoff management. The retrieval step returns material fragments related to a question. Permission decisions and handoff execution are separate business actions. Update frequency does not change what this page's feature does, so it is not developed further here.
Constraints at the retrieval step
A retrieval result reflects how closely document content matches a question. A similarity score does not act as a permission filter and does not distinguish whether a party holds authorization. When a handoff list fragment is recalled, that means the fragment is semantically close to the question. It does not mean the handoff recorded in it has been approved or accepted by the carrier.
Linking knowledge bases requires the same vector model, which is an input condition at the node level. Bases using different vector models cannot be linked together. During implementation, check the vector model identifier first, exclude those that do not match, and record what is actually observed. Do not run cross-model retrieval to compare missed recalls. Semantic retrieval, full-text retrieval, and hybrid retrieval suit different document forms, so identifier and waybill number queries need separate tests from natural-language status questions. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process and is not described as a standalone switch. Reranking requires a model and must be enabled by configuration. Image retrieval requires the corresponding image input and model capability. Optional paths shown in the overall document flow diagram do not mean every request executes all of them.
Referenced content is output as an array, and its length can be 0. The empty-result link still runs through, so downstream design must accept an empty array.
Configuration and implementation checks
The following settings and check items relate to this page's feature, built around a hypothetical drill sample. The sample content is fictional and serves only to illustrate the checking method.
Hypothetical drill sample: A warehouse shipping operator submits a handoff request on 10 June 2025. The request text reads: "Hand over parcels with waybill numbers SF0001 to SF0020 from the 10 June batch handoff list to carrier A's pickup point." The material text available for checking reads: "This batch handoff list contains 20 items, the carrier is A, and the handoff status is pending pickup." A draft sentence also awaits judgment: "This batch of parcels has completed pickup handoff." The retrieval question is "handoff status of the batch containing SF0001."
| Configuration or check item | Hypothetical starting point or value | Reason for the choice and comparison object |
|---|---|---|
| Linked knowledge bases | Link only handoff material bases that use the same vector model | The node requires the same vector model; compare the vector model identifiers of each base for consistency |
| Search mode | Use full-text retrieval for waybill number questions and hybrid retrieval for status questions | Compare recalled fragments for the same question under both modes and observe whether the handoff list text is hit |
| Reference limit | For this drill, set the starting point to a token amount covering one complete handoff list | This value is a trial starting point; adjust after comparing with the actual list text length, and do not treat it as an official default |
| Minimum relevance | Label relevant and irrelevant samples first, then take a trial value between 0 and 1 | This parameter takes effect only under semantic retrieval or reranking; compare filtering results against the labeled samples |
| Question optimization | For "handoff status of the batch containing SF0001," compare the original question with the optimized question | Check whether optimization keeps the waybill number and batch subject, and avoids rewriting the subject |
| Empty-result handling | Use a paper input of [] to check the downstream design | Referenced content length can be 0; check whether downstream treats it as an empty array |
Three common errors
- Treating a retrieval score as a permission decision. The input is a handoff list fragment with a high similarity score. The comparison object is the authorization status of the party the fragment belongs to. The observable result is that score level and authorization status have no correspondence. The error lies in inferring operating permission from recall.
- Inferring completed handoff from a pending-pickup status. The input is the material text "handoff status is pending pickup." The comparison object is the draft sentence "has completed pickup handoff." The observable result is that the two status values do not match, so the draft sentence is not supported.
- Linking bases that use different vector models. The input is the vector model identifiers of two knowledge bases. The comparison object is the node's same-vector-model condition. The observable result is that when the identifiers differ, the base should be excluded and not used in a retrieval comparison.
How to verify the setup
- For the waybill number question "handoff status of the batch containing SF0001," input the full-text retrieval mode and the handoff list text. Compare whether the recalled fragment contains "the carrier is A, and the handoff status is pending pickup." The condition is met when the fragment is recalled and the status field is complete.
- For the draft sentence "this batch of parcels has completed pickup handoff," input the three material sentences and the draft sentence. Compare the values of the same field, "handoff status." The observable result is that the material says pending pickup and the draft says completed, so the draft sentence is judged unsupported.
- For empty-result handling, input the paper empty array
[]. Compare whether the downstream design treats it as an empty reference. The condition is met when the link continues to run and produces no fabricated content. If retrieval returns a non-empty result, record the actual fragment and check it separately, without inferring permission or completed handoff from it.