What the documents look like
Near-expiry markdown handling tracks the lifecycle of shelf inventory rather than weighing individual order items. Store staff record near-expiry batches, markdown approvals, damage write-offs and promotional shelf labels on a daily and shift-based rhythm. Procurement and supplier material therefore holds two kinds of content at once: supplier qualifications, agreements and delivery information on the procurement side, and store-side near-expiry and loss disposal records on the other. Updates arrive frequently, and one batch may produce a near-expiry list, a markdown approval and a damage write-off in sequence, each with its own subject, object and status that need separate checking.
When this material enters a knowledge base, the parsing and chunking step has to keep every passage traceable to its source text and keep a condition statement in the same segment as the object it constrains. Procurement and supplier documents often carry identifiers, qualification names and dates, so full-text retrieval and semantic retrieval each have a role.
Constraints at the document chunking step
According to the official documentation, a knowledge base consists of libraries, collections and data. The smallest search unit is the library; collections only classify data and do not affect search results. One set of data can map to multiple vectors, and the indexes field records the corresponding vector IDs. During retrieval, vectors are recalled first, then the vector IDs are used to fetch original data from the dataset.datas table in MongoDB. When multiple recall paths hit the same set of original data, the results merge and the highest score is kept. Segmentation should keep semantics complete and singular. A leaner index raises precision while sacrificing some retrieval range, and one chunk can carry multiple index entries.
The following is a hypothetical paper exercise, used only to illustrate the checking method.
Hypothetical source excerpt (supplier material): "This page is the annual supply material for Supplier A. The quotation validity period runs from 1 March 2025 to 31 May 2025. Near-expiry markdown disposal for this supplier must be initiated by the store loss administrator and executed after store manager approval." Hypothetical draft sentence to be judged: "Supplier A's quotation remains executable in June 2025, and markdown disposal is decided by the loss administrator alone." The draft sentence has Supplier A as subject, quotation and markdown disposal as objects, and June executability plus unilateral decision as states. The effective range and the decision authority do not match the source. This sample is a paper exercise; nothing was submitted to any system and no approval or execution occurred.
Configuration and implementation checks
The table lists only settings documented in the official material or implementation checks relevant to parsing and chunking on this page. Any numeric value is a hypothetical starting point for this exercise, with the reason for choosing it and the comparison object stated.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Material contains identifiers and natural-language description; semantic and full-text recall both need observation |
| Reference cap | Hypothetical 2000 tokens | The documentation controls reference cap by tokens; the starting point observes whether long segments crowd out other content, compared with a smaller value |
| Minimum relevance | Hypothetical 0.3 | Takes effect only under semantic retrieval or result reranking; observes whether low-relevance fragments are filtered |
| Question optimization | Hypothetically enabled | Material contains follow-up questioning; compare recalled content for the same question after enabling |
| Segmentation method | Segment by document structure | Keeps condition statements in the same segment as the constrained object; compare recalled content before and after splitting |
| Training status check | Check after upload | Upload success and retrieval being updated are separate facts; verify actual training and index state |
Three common errors
- Treating upload success as retrieval being updated. The upload action and the training or index state are different facts. Without checking training status, retrieval may still point at old segments. Compare the upload record with the actual training state and observe whether retrieval returns new content.
- Treating quotation validity period as supply term. In the hypothetical sample, quotation validity runs from 1 March to 31 May 2025, and that field constrains only the quotation; no supply term is stated. The draft sentence saying June remains executable is unsupported. Compare the value and period of the same field to confirm the same subject, the same period and the same field.
- Treating retrieval or a form as authorized execution. Markdown disposal must be initiated by the loss administrator and approved by the store manager. A retrieval hit on the approval description does not mean approval is complete. Compare the request status with the approval record; a pending status does not show that authorization was granted.
How to verify the setup
The three checks below use the hypothetical source excerpt and draft sentence above, stating input, comparison and observation for each. No actual test is claimed.
- Input the hypothetical source excerpt and the draft sentence, and compare the quotation validity field with the June executability stated in the draft. If the returned segment contains both the validity clause and the markdown approval authority, and the draft sentence disagrees with the source on period and authority, the draft sentence is unsupported and the evidence gap should be recorded.
- Input the markdown disposal portion of the draft sentence and compare the initiating and approving parties stated in the source. If the returned content shows the initiator is the loss administrator and the approver is the store manager, while the draft says the decision is unilateral, the wording does not match the source. Keep the source clause and mark the item as pending verification.
- Input a query containing only the supplier name, and observe whether the returned segments include that supplier's qualification, quotation and disposal clauses. If the returned content is spread across multiple segments with conditions and objects separated, adjust the segmentation method and compare again. If the return is empty, a clearly stated paper input of an empty set can be used to check downstream design, without inferring that the rule necessarily returns an empty result.
Retrieval does not carry authorization or entity isolation duties. Minimum relevance is a filtering threshold and does not exclude unauthorized records. Libraries built on different vector models cannot be associated together; verify identifiers first, exclude those that do not match, and record the actual observation.