What the documents look like
A third-party overseas warehouse can serve several sales channels, and the seller allocates stock across them. The working documents are replenishment plans, inbound appointment records, inventory age tables and last-mile delivery orders. These come from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems and carrier bills, with weekly or daily updates. In order fulfilment and after-sales answering, citation tracing means each quoted replenishment quantity, appointment window, inventory-age charge or delivery status can be matched back to the wording of a specific document. When material is revised, whether an earlier plan or earlier appointment may still serve as a citation basis depends on checking the effective version.
Constraints at the citation tracing step
The knowledge base search node may associate one or more knowledge bases that use the same vector model. Its output is a citation array whose length may be zero; the output path still runs when nothing is found. The citation cap controls, in tokens, the maximum content cited per search. The minimum relevance value is a 0–1 filter threshold that applies only in semantic retrieval or when result reranking is used; it filters low-relevance results and does not perform permission or entity isolation. Hybrid retrieval merges multiple recall paths through RRF, and a rerank model returns a 0–1 score used for filtering. Semantic retrieval, full-text retrieval, image retrieval, question optimisation and reranking are optional paths; whether each runs depends on the input and settings, and not every request executes all of them.
Hypothetical paper exercise sample: the source wording is "The replenishment plan dated 10 March 2025 lists SKU A-100 for the US West warehouse at 600 units, with an inbound appointment window of 18 March 2025, 09:00 to 11:00. The replenishment plan dated 12 March 2025 adjusts the SKU A-100 replenishment quantity to 450 units; the inbound appointment window is unchanged." A draft sentence to be assessed is "The SKU A-100 replenishment quantity at the US West warehouse is 600 units, with an inbound appointment window of 18 March 2025, 09:00 to 11:00."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only the replenishment plan base and inbound appointment base confirmed to use the same vector model | Check the base identifier and vector model identifier, exclude bases whose models differ, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Replenishment plans contain identifiers and dates; hybrid retrieval uses semantic recall and full-text recall together and merges them through RRF |
| Citation cap | Set to 2000 tokens for this exercise | This is a hypothetical starting point; compare it with the length of the sample source text and observe whether the citation array contains the complete plan sentence |
| Minimum relevance | Set to 0.5 for this exercise | This is a hypothetical starting point; compare it with the rerank score and observe whether candidates below 0.5 are filtered |
| Result reranking | Enabled | Rerank text candidates when there are many, and filter using the rerank score |
| Reply format | Keep the default | After-sales answering needs natural-language replies, and no downstream step reads a fixed structure |
Three common errors
- Treating minimum relevance as a permission or entity filter. Observable error: assuming that raising minimum relevance excludes other channels or unauthorised records. Check: input the sample source text, compare whether the returned citations are decided by similarity alone, observe whether an unrelated but semantically close record can still be recalled, and confirm that permission isolation requires a separate system implementation.
- Treating retrieval output as authorised execution. Observable error: taking the replenishment quantity in the citation array as approved for execution. Check: compare the draft sentence with the source text, observe that the 12 March 2025 version already adjusted the quantity to 450 units, confirm that 600 units appears only in the earlier version, and confirm that execution depends on review against the effective version.
- Treating an empty citation array as a workflow fault. Observable error: declaring the workflow failed when the citation length is zero. Check: use the explicit paper input [] to carry the downstream design, observe whether the output path runs, and confirm that an empty array does not equal a terminating error.
How to verify the setup
- Input the two replenishment plan versions from the hypothetical sample, compare the draft sentence "the replenishment quantity is 600 units" with the 450 units recorded in the 12 March 2025 version, observe that the same subject, same period and same field carry mutually exclusive values, and judge that the draft sentence is not supported by the effective version.
- Input the inbound appointment sentence from the hypothetical sample, compare the draft sentence "the appointment window is 18 March 2025, 09:00 to 11:00" with the source text, observe that neither plan version changed that window, and judge that this part matches the source text.
- Input a short question containing only "replenishment", compare the returned citations with the sample source text, observe that the citation array may be empty or may contain only low-relevance fragments, and record the actual observation without presuming a first hit or full recall.