What the documents look like
Third-party overseas warehouses can serve several sales channels at once, and the seller allocates stock across those channels. This differs from a single-channel replenishment arrangement, and it means a product question may need to draw on records that different parties maintain. The working material for this category arrives from several systems: the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems, and carrier billing. The recurring documents are replenishment plans, inbound appointment records, inventory age tables, and last-mile delivery orders. Their update frequency is weekly or daily, and that cadence affects how current a cited record is, not whether the citation feature works.
For product data and selection questions, the model numbers, quantities, appointment statuses, and fee definitions inside those documents serve as citation sources. A useful answer traces each claim back to the wording of the source record. Stock allocation and inbound appointment execution sit with different parties, so a retrieval result supplies a pointer to source text and nothing more. The three most time-consuming steps reported for this category are forecasting sales, booking inbound appointments, and reconciling inventory age fees; each of these produces records that later questions may cite.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it returns citations as an array whose length can be zero. An empty array means the search path completed and produced no citation content. That output then passes to the answering stage, where each statement is checked against the cited source text. The citation limit is expressed in tokens, not in a count of items. The minimum relevance value takes effect only under semantic retrieval or result reranking. Semantic recall and relevance scores do not perform permission filtering or entity isolation.
Hybrid retrieval merges multiple result paths through RRF, and reranking requires a rerank model and runs according to its setting. Image retrieval requires an image input together with the relevant model capability. Knowledge bases built on different vector models cannot be associated together; the identifiers are checked first, non-matching bases are excluded, and the actual observation is recorded.
Hypothetical exercise sample: a cross-border seller holds a replenishment plan whose text reads, "On 10 March 2025, SKU A had 120 units of available stock at the US West overseas warehouse, with a planned replenishment of 300 units before 5 April, and the inbound appointment status is pending review." A draft sentence reads, "This SKU's replenishment request has been approved and will arrive at the warehouse before 5 April." The source states only that the appointment status is pending review; it does not state an approval. The draft sentence shares the same subject and the same period but carries a different value for the same field, so it is unsupported.
Configuration and implementation checks
The settings and checks below relate to this page's sample. The numeric values are starting points for this hypothetical exercise.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only knowledge bases sharing the same vector model as the sample material | Check the knowledge base identifier and vector model identifier, and exclude bases on different vector models |
| Search mode | Hybrid retrieval | The sample contains an SKU number, so full-text retrieval can match the number while semantic retrieval adds meaning |
| Citation limit | 2000 tokens as the hypothetical starting point | The sample source text is roughly 80 characters; compare against single chunk length before adjusting |
| Minimum relevance | 0.5 as the hypothetical starting point | Takes effect only under semantic retrieval or reranking; compare against rerank scores before adjusting |
| Result reranking | Enabled, provided a rerank model is configured | Check the model identifier and enablement state; do not enable when no model exists |
| Citation content output | Accept the array, allowing length zero | When the array is empty, record the gap rather than treating it as an error |
Three common errors
- Treating retrieval output as an authorization conclusion. The input is the source text and the draft sentence. The comparison is the field value: the source says pending review, the draft says approved. The observable result is that the two values conflict for the same subject, period, and field, so the draft sentence is unsupported and the evidence gap remains.
- Treating the minimum relevance value as a permission filter. The input is the minimum relevance setting and the knowledge base scope. The comparison is whether that setting is used to exclude records the requester may not access. The observable result is that the setting affects relevance filtering only and does not perform permission isolation.
- Treating an empty citation array as workflow termination. The input is a paper array [] with no elements. The comparison is whether the downstream node continues to run. The observable result is that the downstream design should accept an empty array, and business failure is not concluded from it.
How to verify the setup
- Citation tracing check. The input is the hypothetical source text and the draft sentence given above. The comparison is the two field values, pending review against approved. The observable result is that when the values conflict, the draft sentence is unsupported and the evidence gap is preserved.
- Empty citation handling check. The input is the paper array []. The comparison is whether the downstream answering node keeps running. The observable result is that the path completes without citation content; no claim is made that retrieval always returns an empty array.
- Cross-model association check. The input is the vector model identifiers of two knowledge bases. The comparison is whether the identifiers match. The observable result is that when the identifiers differ, the association is excluded, and no cross-model retrieval is run to compare missed recall.