What the documents look like
Receiving and put-away covers the first physical-and-record confirmation that goods have entered the warehouse. The working material comes from a WMS, ERP purchasing records, supplier delivery notes, and barcode scanning. The everyday documents are the purchase arrival note, the receiving inspection sheet, the put-away task, and the discrepancy report. These documents update per batch and per day.
For order fulfilment and after-sales answers, the records in front of a retrieval step are not uniform. Arrival quantity, inspection conclusion, assigned storage location, and discrepancy handling status sit in different documents, and a single batch may be spread across all of them. A question about one batch needs the material for that same batch. Retrieval returns reference content only; it does not decide authorization.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. At the retrieval step, confirm each knowledge base identifier first, exclude those whose vector model does not match, and record what was actually observed. Do not run a cross-model search in order to compare missed recalls.
The node outputs references as an array, and the length can be 0. When nothing is found, the output path still completes. Semantic retrieval depends on vector similarity; full-text retrieval depends on keyword hits; hybrid retrieval merges the two paths through RRF, and RRF is a fusion process. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold. The reference cap is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion, and it adds one model call.
The same-vector-model requirement for linked knowledge bases does not extend to business registration sites or other external systems.
Hypothetical paper exercise: an e-commerce warehouse receiving team processes purchase arrival note DH-20250310-07 dated 10 March 2025. The supplier delivery note lists 500 units of screw M4×12. The receiving inspection sheet records 480 units actually received. The discrepancy report draft states "short by 20 units, pending supplier confirmation." A draft sentence awaiting judgment reads: "The 20-unit shortage in this batch has been confirmed by the supplier." The retrieval scope is limited to the knowledge base holding the receiving inspection sheet and discrepancy report for this batch. A person compares the source wording before deciding whether the draft sentence holds.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the base holding the receiving inspection sheet and discrepancy report | Check each base's vector model identifier first, exclude mismatches, then link |
| Search mode | Hybrid retrieval | Run the same test questions against semantic, full-text, and hybrid modes, then compare recalled content |
| Reference cap | Assumed starting point of 2000 tokens | Test content completeness with long and short real material, compare against smaller values, then adjust |
| Minimum relevance | Assumed starting point of 0.4 | Label relevant and irrelevant samples first, compare filtering against higher and lower thresholds |
| Question optimization | Enabled for multi-turn follow-ups | Compare the original question with the optimized question, check whether the subject was altered |
| Result reranking | Enabled according to actual model capability | Enable when a rerank model exists and candidates are numerous, compare ordering before and after |
Three common errors
- Treating minimum relevance as a permission or entity filter. Minimum relevance is a filter threshold that takes effect only under semantic retrieval or when result reranking is used. A similarity score does not distinguish business ownership, and retrieval results do not constitute authorization. Check: submit the same question, compare cited content under different relevance thresholds, and observe that a threshold change affects relevance ordering only, not the accessible range.
- Describing RRF as a switch that can be turned on. RRF is the fusion process for multiple result paths in hybrid retrieval and has no corresponding switch. Check: inspect the configuration interface for such a switch; if none exists, record RRF as a fusion process and do not configure it as a standalone option.
- Writing a pending status as completed confirmation. In the hypothetical sample, the discrepancy report states "pending supplier confirmation," while the draft sentence says "confirmed by the supplier." Both share the same subject, batch, and field, and the two values are mutually exclusive. Check: input the discrepancy report source text and the draft sentence, compare the confirmation status field, and observe that the draft sentence does not hold while the source remains pending.
How to verify the setup
- Input the discrepancy report source text from the hypothetical sample and the draft sentence "The 20-unit shortage in this batch has been confirmed by the supplier." Compare the source confirmation status with the draft wording. The source reads "pending supplier confirmation," and the draft states confirmed; the two conflict, so the draft sentence is not supported. Expected observation: the comparison is marked as pending verification.
- Input the purchase arrival note and receiving inspection sheet from the hypothetical sample. Compare the arrival quantity of 500 units with the actual receipt of 480 units. These are different fields; the 20-unit difference holds, but whether the supplier confirmed still requires support from the discrepancy report source text. Expected observation: the quantity difference can be checked, and the confirmation status retains an evidence gap.
- Input an empty reference array and check the downstream path design. The node output length can be 0; when the result is empty, the output path still completes, and downstream handling must treat empty references accordingly. Expected observation: after receiving an empty array, downstream generates no cited content, and this result does not indicate that the retrieval rule always returns an empty array.