What the documents look like
Complaints and official replies face platforms or regulators, so the file set carries formal evidence and rectification wording. The main sources named in the brief are the 12315 platform, platform complaint centres, customer service chat, and the order system. Everyday documents are complaint tickets, regulator inquiry letters, order evidence, and rectification reports. Typical roles are customer service supervisors, legal compliance staff, and store owners. Errors can lead to administrative investigation, platform penalties, and brand-level public reaction. Updates are irregular and per case.
Cross-border material and multilingual retrieval means one case may hold a Chinese ticket, foreign-language order proof, and a regulator inquiry letter at the same time. The retrieval step has to map material in different languages into one knowledge base. Because updates are per case, retrieval settings are re-checked whenever case material changes.
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. The only action available is to check the vector model identifier of each base, exclude those that do not match, and record what was observed. No conclusion follows about how a combined link would behave.
Semantic retrieval supports cross-language understanding, so a Chinese question can match an English knowledge point. Its precision depends on model training quality and is affected by keywords and sentence completeness. Full-text retrieval suits precise lookups such as numbers, model codes, proper nouns, and error codes. Hybrid retrieval merges semantic and full-text results through RRF. RRF is a fusion process inside hybrid retrieval, not a switch that can be turned on. The citation cap is measured in tokens. The minimum relevance value is a 0–1 filter threshold and applies only under semantic retrieval or when result reranking is used. Question optimisation handles coreference removal and question expansion and adds one model call before retrieval. Retrieval itself does not perform permission or entity isolation. Finding content through retrieval does not mean authorisation has been granted.
A hypothetical paper exercise sample: a store receives a regulator inquiry letter. The letter text reads: "Please submit a written explanation and evidentiary materials on the refund handling of order A-2024-0317 within ten working days of receiving this letter." A customer service supervisor drafts the handover sentence: "This inquiry letter has been approved by the store's legal team and can be submitted externally as is." In this sample, the letter states a submission deadline and material requirements. It does not state legal approval or external submission authorisation. When retrieving this letter, the search term used is "order A-2024-0317 refund regulator inquiry letter". The observation is whether the recalled content contains the letter text above. If the recalled result contains only the order refund record and not the letter text, the evidence gap is recorded and no approval is concluded.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link one knowledge base holding complaint tickets and inquiry letters | Check the vector model identifier of each base; exclude any that do not match |
| Search mode | Test semantic retrieval and full-text retrieval separately first | Use the same test question to compare recalled content; do not assume one mode fits all material |
| Citation cap | Set 2000 tokens for this exercise | Test with short and long material to see whether content stays complete, then adjust |
| Minimum relevance | Set 0.3 for this exercise | Label relevant and irrelevant samples first, then compare recall differences as the threshold changes |
| Question optimisation | Enable only for multi-turn follow-ups or short questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable when candidate results are numerous | Confirm the rerank model is available; compare ranking before and after reranking |
Three common errors
- Treating a retrieval hit as authorisation. A search for "order A-2024-0317 refund" returns a refund record, and the draft is then treated as submittable externally. Check: compare the inquiry letter text for a stated authorising party and authorisation scope; observe whether the draft sentence "approved by the legal team" has matching source text; if there is no support, record the item as pending verification.
- Linking knowledge bases across vector models. Two bases using different vector models are linked together, and a conclusion is then drawn about missed recall or redundancy. Check: compare the vector model identifier of each base, exclude those that do not match, record the actual observation, and do not infer runtime behaviour.
- Writing a pending state as a completed state. In the draft sentence "this inquiry letter has been approved by the store's legal team", the letter text states only a submission deadline and does not state approval. Check: compare the same subject, the same period, and the same field; if the approval field has no source support, the item is judged unsupported.
How to verify the setup
- For the sample inquiry letter, enter the search term "order A-2024-0317 refund regulator inquiry letter" and compare whether the recalled content contains the letter sentence "submit a written explanation and evidentiary materials within ten working days". If the recalled content contains that sentence, retrieval covers that evidence. If it does not, record the gap and do not conclude that approval was granted.
- For the draft sentence "this inquiry letter has been approved by the store's legal team and can be submitted externally as is", compare against the inquiry letter text to see whether legal approval and external submission authorisation are stated. The source text does not state them, so the draft sentence is judged unsupported and the evidence gap is retained.
- For empty-result handling, use the paper input [] to check the downstream design and observe whether the output link still runs when the citation content length is 0. Do not claim that a fictional rule search necessarily returns [].
For the empty-result case, the documented output is an array of citations whose length can be 0, meaning the output link still runs even when no content is found. This is a documented node behaviour and does not depend on any assumed scheduled trigger or hidden control. Forms collect values; business verification, approval, database writes, and external synchronisation require their own downstream implementation and are not performed by retrieval.