What the documents look like
Warehouse picking, checking and packing handles outbound orders rather than supplier arrivals. The operational data comes from the WMS, the order system, the electronic waybill platform and weighing equipment. The documents handled each day include wave picking lists, checking records, waybills and packaging consumable sheets. Procurement and supplier records sit beside these outbound documents as cross-stage reference material, used to check consumable ownership, supplier names and purchase records. Their update rhythm differs from that of outbound paperwork.
This page addresses one task only: knowledge-base retrieval and recall for procurement and supplier records inside the picking, checking and packing category. Retrieval locates reference text. It does not approve a purchase, confirm a supplier or authorise a handover. A record that appears in the retrieved content remains a record, and its stated status is the only status the material supports.
Because retrieval questions in this setting often mix order numbers, consumable codes and supplier abbreviations, semantic retrieval and full-text retrieval behave differently across record types. A consumable code is a precise string suited to keyword matching. A supplier short name written in natural language may match better through vector similarity. The two paths are combined rather than treated as substitutes.
Constraints at the retrieval step
The knowledge-base search node can attach one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; collections inside it serve classification management and do not change search behaviour. Attaching bases with different vector model identifiers is outside the documented condition, so the identifiers are checked first and mismatched bases are excluded before any search runs.
Semantic retrieval computes distance between vectors and depends on model training quality. Full-text retrieval depends on keyword hits and suits codes, model numbers, proper nouns and error codes. Hybrid retrieval uses both and merges the two result sets through RRF. A ReRank model reorders candidate text and returns a 0–1 score representing relevance. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The citation limit is measured in tokens, not in a fixed item count.
Retrieval results are not authorisation. A retrieved purchase record does not mean the purchase was approved, and a retrieved supplier card does not mean the supplier is active. When no content is found, the reference output is an array whose length can be zero, and the downstream chain still runs. The optional paths shown in the overall retrieval flow are not executed on every request: image retrieval requires an image input and matching model capability, and reranking requires a model that is enabled in the settings.
Hypothetical paper exercise sample
For checking purposes, the source material reads: "Packaging consumable sheet HC-0712 states: consumable name five-layer corrugated carton, specification 60×40×40 cm, quantity drawn 200 pieces, draw date 3 June 2025, drawing warehouse East China Warehouse No. 2." A second source reads: "Supplier record card GYS-033 states: supplier name Antai Packaging Materials Co., Ltd., supplied category corrugated cartons, cooperation status pending review, record update date 28 May 2025."
The draft sentence to be judged is: "Antai Packaging Materials Co., Ltd. has passed review and may supply five-layer corrugated cartons to East China Warehouse No. 2." This draft carries a subject, an object and a status, and each element is compared against the two source texts above.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Attached knowledge bases | Attach only the procurement and supplier records base | The vector model identifier of this base and of the outbound document base are compared first; attach only after the identifiers match |
| Search mode | Run hybrid retrieval first | The sample contains code HC-0712 and a supplier short name; compare semantic hits against full-text hits |
| Citation limit | Hypothetical starting point 2000 tokens | The sample holds two source texts; compare content completeness against a 1000-token setting |
| Minimum relevance | Hypothetical starting point 0.4 | Label relevant and irrelevant samples first, then compare filtering against 0.3 and 0.5 |
| Question optimisation | Enable first | The sample contains references and abbreviations; compare the original question against the optimised question to check whether the subject was altered |
| Rerank | Enable first | When candidates are numerous, compare rerank score ordering against vector score ordering |
The numeric values above are starting points for this hypothetical exercise and are not official defaults.
Three common errors
- Treating a retrieval hit as completed authorisation. Input the draft sentence and the supplier record card. Compare the cooperation status value "pending review" against the draft phrase "has passed review". Observe whether the two occupy the same field with mutually exclusive values. Pending review does not constitute passed review.
- Treating a quotation validity period as a supply term. If the record card carries a separate quotation validity field, compare that field against the supply term field and observe whether they are the same field. A quotation validity period does not state a supply term.
- Attaching knowledge bases that use different vector models. Input the vector model identifiers of the two bases, compare whether the identifiers match, and observe whether a mismatched base is excluded. Cross-model retrieval is not executed in order to compare missed recalls.
How to verify the setup
- Input the hypothetical source texts and the draft sentence. Compare the cooperation status field value against the draft claim. Observe whether the draft sentence obtains support from the source text. If the source text states only "pending review", the draft phrase "has passed review" is unsupported, and the evidence gap is recorded.
- Input the drawing warehouse from consumable sheet HC-0712 and the supply object named in the draft sentence. Compare the warehouse identifier against the supplier name, and observe whether subject and object align. If they do not align, record the item as pending verification rather than declaring a contradiction.
- Input an empty-result scenario, using a paper input of an empty array, and check the downstream design. Observe whether the output chain still runs. No claim is made that a fictional rule search must return an empty array.
Across all three checks, a determination requires the same subject, the same period and the same field. Only a single field carrying explicitly mutually exclusive values counts as a contradiction. A value that is absent, or evidence that shows only a related stage, counts as unsupported or pending verification. Pending review does not demonstrate a lack of operating permission. Signing, reaching cooperation and contract effectiveness are three distinct facts. A quotation validity period does not state a supply term. No authorisation or effectiveness clause is invented to obtain a definite answer, and the evidence gap is left open.