What the documents look like
Channel diversion reviews for order fulfilment and after-sales questions work with anti-counterfeit code records, e-commerce platform price screenshots, distributor shipment records, market inspection photographs and diversion complaint forms. These materials span regions and batches, so one shipment can correspond to several distributor shipment records and several low-price platform listings. Updates arrive weekly and irregularly, which means retrieval faces complaint forms, penalty notices and investigation photographs that are not on a single version. The retrieval and recall step exists to pull fragments relevant to the current complaint form or after-sales request so that a person can review them. Retrieval does not decide whether diversion occurred and does not execute a penalty.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the vector model identifier must be checked first and non-matching bases excluded, and retrieval across models is not performed in order to compare missed recalls. Search mode is compared per test question across semantic, full-text and hybrid retrieval; no single mode is presumed suitable for all material. The citation limit is measured in tokens, so content completeness is tested with materials of different lengths before the limit is adjusted. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimisation performs coreference resolution and question expansion, so the original question and the optimised question are compared to check whether the subject was altered. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch. Image retrieval requires a corresponding image input and model capability, and reranking requires a model and must be enabled according to its setting. Retrieval and form collection gather values; they do not constitute authorisation, approval or a database write.
Configuration and implementation checks
The configuration below is built around a hypothetical paper exercise. The sample content is constructed on paper and does not represent a real customer or a business event that has occurred.
Hypothetical paper exercise: a brand channel control team receives a diversion complaint form whose original text reads: "Complaint subject: East China distributor A; anti-counterfeit code involved: FC-2024-8871; complaint reason: the product corresponding to this code is sold on e-commerce shop B below the recommended retail price; complaint date: 3 June 2024; attachments: one platform price screenshot, two inspection photographs." A draft sentence awaiting determination reads: "Distributor A has completed the diversion penalty." The material scope to be checked includes this complaint form, the corresponding anti-counterfeit code record and the draft penalty notice. The draft penalty notice is at the pending-review stage and is unsigned, so it does not constitute a completed penalty.
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select the diversion complaint base and the anti-counterfeit code record base that use the same vector model | Check first whether the two vector model identifiers match, exclude non-matching ones, and do not perform retrieval across models |
| Search mode | Use full-text retrieval for "FC-2024-8871" and hybrid retrieval for "low-price sale complaint" | Compare full-text hits for code-type questions and compare semantic plus full-text fusion results for natural-language questions |
| Citation limit | Start at 1024 tokens | Compare one short complaint form and one long inspection report to see whether recalled content is truncated, then adjust |
| Minimum relevance | Start at 0.5 | Label relevant and irrelevant samples first, then compare whether the complaint subject and anti-counterfeit code survive filtering |
| Question optimisation | Enable for multi-turn follow-ups, disable for single-turn explicit questions | Compare the original question and the optimised question to check whether references such as "this code" or "this distributor" were altered |
| Result reranking | Enable ReRank when candidates are numerous | Compare the 0-1 score before and after reranking to confirm whether the complaint form and anti-counterfeit code record move forward |
Three common errors
- Treating retrieval recall as authorised execution. The input is the text "distributor A sold below price" in the complaint form, and the comparison object is the pending-review status of the draft penalty notice. Treating recalled content as an effective penalty confuses retrieval with approval. The observable result is that the draft penalty notice remains at pending review and is unsigned.
- Associating knowledge bases across vector models. The input is the vector model identifier of the diversion complaint base and of the anti-counterfeit code record base, and the two are compared for consistency. Attempting joint association when the identifiers differ violates the same-vector-model requirement. The observable result is that the two model identifiers differ, the non-matching base should be excluded, and the observation should be recorded.
- Treating minimum relevance as an authorisation or entity filter. The input is a minimum relevance of 0.5 and one irrelevant distributor record, and the comparison is whether that record is filtered out for low similarity. Treating the setting as capable of excluding unauthorised distributors misreads the nature of the filter. The observable result is that filtering acts only on the similarity score and does not determine a distributor's authorisation scope.
How to verify the setup
- Verify full-text recall of the anti-counterfeit code. Input "FC-2024-8871" from the hypothetical complaint form and compare whether the full-text retrieval result contains the record corresponding to that code. The condition is met when the recalled content contains the anti-counterfeit code and the complaint subject. If neither appears, record it as not recalled and check whether the data has completed training and how tokenisation is configured.
- Verify consistency between the draft sentence and the source text. Input the draft sentence "Distributor A has completed the diversion penalty" and compare it against the status of the draft penalty notice. The condition is met when the draft remains at pending review and unsigned, so the draft sentence is unsupported. If the draft has been signed, the signing time and authorisation scope must be checked separately.
- Verify downstream handling of empty citation output. Input the paper-constructed empty result array [] and compare whether the downstream node continues as designed with a citation length of 0. The condition is met when the output path is not interrupted and the downstream node does not treat the empty result as recalled content. This verification does not claim that a fictional rule retrieval necessarily returns []; it checks the design for handling an empty result.