What the documents look like
Receiving and putaway is the first physical-and-record confirmation that goods have entered the warehouse. The working materials come from WMS, ERP purchasing, supplier delivery notes and barcode scanning, and the daily documents are purchase arrival notes, receiving inspection sheets, putaway tasks and discrepancy reports.
Contract and transaction material comparison places the parties, objects, quantities and states recorded in the purchase arrival note and the receiving inspection sheet onto the same fields. Retrieval therefore has to cover two kinds of content in one batch. Arrival note numbers, purchase order numbers and barcodes are identifier-type content, and full-text retrieval depends on keyword hits for these. Inspection conclusions and discrepancy reasons are natural-language content, and semantic retrieval depends on vector similarity for these. Because both kinds sit in the same batch of materials, the search mode and the reference limit directly affect whether a complete field set can be retrieved for comparison.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. When identifiers differ, the operator first checks the identifiers, excludes the non-matching base, and records the actual observation; a cross-model search is not run in order to compare results.
The reference limit is measured in tokens. Arrival note fragments and discrepancy report fragments differ greatly in length within one batch, so a token-based limit is steadier than a count of items. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold and does not carry out permission or entity isolation; a retrieval result does not constitute authorization.
Hybrid retrieval merges the semantic and full-text paths through RRF. RRF is a fusion process and is not written as an on/off switch. The image retrieval and reranking paths shown in the overall document flow are optional: they require the corresponding image input and model capability and must be enabled according to settings, and they do not run on every request. On an empty result the reference content is output as an array and its length can be 0, and the downstream chain still proceeds.
Configuration and implementation checks
The table below is organised around a hypothetical drill sample. All values are starting points for this hypothetical trial, used to compare against labelled samples before adjustment, and they are not official default values.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the receiving material base that uses one vector model | Check each base's vector model identifier first, exclude non-matching ones, and do not run a cross-model search |
| Search mode | Hybrid retrieval | The sample contains an arrival note number and an inspection description; compare the single-path full-text and semantic results with the hybrid result for completeness of hits |
| Reference limit | Hypothetically set to 800 tokens | The sample source text is about 300 characters; observe at this value whether the arrival note number and the inspection conclusion both enter the reference, and compare with 400 and 1200 |
| Minimum relevance | Hypothetically set to 0.4 | Label relevant and irrelevant fragments in the sample first, observe whether 0.4 filters irrelevant fragments while retaining the inspection conclusion, and compare with 0.3 and 0.5 |
| Question optimisation | Hypothetically enabled for this trial | The sample draft sentence contains a reference; compare the original question with the optimised question and check that the subject was not altered |
| Result reranking | Hypothetically enabled for this trial | The sample has many candidates; compare the rerank score with the vector score for ordering, and observe whether the inspection conclusion moves forward |
Hypothetical drill sample: the source text supplied for checking reads, "Purchase arrival note PO-2024-0517 states that Supplier A delivered 120 units of Type A storage boxes on 17 May 2024; receiving inspection sheet YS-2024-0517 records 118 units actually received, a discrepancy of 2 units, and the discrepancy reason field is blank; putaway task SJ-2024-0517 has the status awaiting bin assignment." The draft sentence to be assessed reads, "This batch has passed inspection and completed putaway." That sentence touches two fields, the inspection conclusion and the putaway state. The source text carries only the received quantity and the awaiting-assignment status; passing inspection and completed putaway are not stated.
Three common errors
- Treating a retrieval hit as authorization or as physical-and-record confirmation. The input is the sample source text above together with the draft sentence; the comparison is between the source text's "awaiting bin assignment" and the draft sentence's "completed putaway." Accepting the draft sentence directly is the error, and the observable result is that an unsupported state is written as completed.
- Associating knowledge bases built on different vector models. The input is two receiving material bases whose vector model identifiers differ; the comparison is between the two identifiers. Associating them directly and comparing recalled results is the error. The correct action is to check the identifiers, exclude the non-matching base, and record the observation.
- Treating minimum relevance as a permission or entity filter. The input is a retrieval request containing an unrelated supplier fragment; the comparison is between the threshold setting and whether the unrelated fragment is filtered out. Concluding from this that unrelated supplier records were excluded by permission is the error; relevance does not perform entity isolation.
How to verify the setup
- Input the sample source text and the draft sentence above. Compare the draft sentence's "passed inspection" and "completed putaway" with the source text's "discrepancy reason field is blank" and "awaiting bin assignment." The observable result is that the draft sentence is judged unsupported, and both passing inspection and completed putaway remain to be verified.
- Input the arrival note number PO-2024-0517 from the sample source text. Compare the reference content from full-text retrieval with that from hybrid retrieval. The observable result is that the reference contains both that number and the received quantity of 118 units; if only a descriptive fragment appears, the condition is not met.
- Input a query unrelated to receiving materials. Compare the returned reference content with the expected empty array. The observable result is that the downstream chain still proceeds when the reference length is 0; this result does not mean that any invented rule necessarily returns an empty array.