What the documents look like
Parcel label and handoff work sits at the point where a parcel leaves the warehouse. Responsibility questions at that point depend on carrier scan records. Operating reports and exception reviews cross-check four kinds of material: electronic waybills, handoff lists, pickup scan records, and exception item registers. The parties, timestamps and status fields across these materials frequently disagree, so retrieval can mix data from different batches or different carriers into one result set.
Knowledge base retrieval addresses a narrower question: given a review question, return the relevant passages from existing materials so a person can compare them. Retrieval does not decide responsibility and does not replace an authorization or approval step. The knowledge base search node can associate one or more knowledge bases that use the same vector model. The node outputs cited content as an array, and the length can be zero, meaning the output path still runs when nothing is found.
Constraints at the retrieval step
A hypothetical drill sample is used below. It is a paper exercise only and describes no real customer, carrier or shipment.
Warehouse shipping group, 15 April 2025, exception item review. Material text one: "12 April handoff list item 7, waybill SF1234567890, delivery address Hangzhou Yuhang District, pickup scan record shows 12 April 20:14 picked up." Material text two: "Exception item register, 12 April entry: same waybill number, registered reason delivery address does not match the label, status pending verification." Draft sentence to be judged: "This waybill was picked up by the carrier on 12 April, so the exception registration should be closed."
Within retrieval, semantic search relies on vector distance, full-text search relies on keyword hits, and hybrid search merges the two result sets through RRF. The minimum relevance value takes effect only under semantic search or when result reranking is used. The citation cap is measured in tokens. Question optimization handles reference resolution and question expansion. Image retrieval requires image input and model capability, and reranking requires a model and must be enabled in the settings.
The pending status in the register and the pickup scan record describe different fields. A scan record showing pickup does not close a pending exception entry. Closing it is a business action that needs a separate downstream process, not a retrieval output.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One trained parcel handoff knowledge base only | This base uses the same vector model as the material under review; compare vector model identifiers across bases and exclude any that differ |
| Search mode | Hybrid search | The sample contains an exact waybill number and natural language description; compare semantic and full-text results separately, then observe whether the RRF-merged candidates cover both field types |
| Citation cap | Hypothetical starting value of 800 tokens | The sample material is short, and 800 can hold two to three complete passages; compare with 400 and observe whether the passage containing the waybill number is truncated |
| Minimum relevance | Hypothetical starting value of 0.3 | First mark relevant and irrelevant passages by hand, then compare 0.2 and 0.4 to see whether relevant passages remain and irrelevant passages decrease |
| Question optimization | Enabled | The draft sentence contains "this waybill", so reference resolution is needed; compare the optimized query with the original and check whether the subject was changed incorrectly |
| Result reranking | Enabled | When candidates are numerous, use the rerank score for filtering; compare ordering before and after reranking and observe the position of the passage containing the waybill number |
Three common errors
- Treating a retrieval hit as an authorization to act. In the sample, the exception register status is pending verification. Finding a pickup scan record does not mean the exception is closed or responsibility is assigned. The retrieval output is cited content only; the business judgment needs manual comparison and a later process.
- Associating knowledge bases that use different vector models. The node requires associated knowledge bases to use the same vector model. When two bases carry different vector model identifiers, check the identifiers first and exclude the mismatched one. Do not run a cross-model retrieval to compare missed recalls.
- Treating minimum relevance as an entity or permission filter. Minimum relevance is a similarity threshold. It does not exclude unrelated businesses or unauthorized records. The retrieval scope is set by the associated knowledge bases, and permission isolation needs a separate system implementation.
How to verify the setup
- Enter the sample material text and the draft sentence, then compare whether the cited content returned by retrieval contains the passage with waybill SF1234567890. Expected observation: the cited content includes that waybill number and the "picked up" field. If the cited content is an empty array, record the empty result and check whether the associated knowledge base has completed training.
- Enter the draft sentence and compare the query terms before and after question optimization. Expected observation: the optimized query keeps the waybill number that "this waybill" refers to, and the subject is not replaced with another batch. If the subject changes after optimization, record that optimization result and check it.
- Enter a short query containing only the waybill number and compare the returns from full-text search and semantic search. Expected observation: full-text search hits the passage containing that waybill number, while semantic search may return semantically close passages with a different waybill number. Record both result sets and do not assume in advance which path will hit.