What the documents look like
Channel diversion review concerns regional and price order. Its materials are diversion complaint forms, anti-counterfeit code records, penalty notices and investigation photographs. These materials enter the knowledge base to support retrieval and recall when product and marketing copy is drafted. A complaint form typically carries the complained party, the region involved, the product object and the acceptance status. An anti-counterfeit code record links a goods code to a query time. A penalty notice records a handling decision and its effective status. An investigation photograph provides on-site or product-appearance information. The weekly and irregular update rhythm means retrieval results reflect the material state at the time of ingestion, and copy drafting is limited to data that has been ingested and has completed training. When the materials hold only a complaint acceptance record, copy stating that a penalty has been executed is not supported. When a photograph has no image description or image vector index generated at ingestion, the image content cannot participate in retrieval.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together; the vector model identifier is checked first, non-matching bases are excluded, and no cross-model retrieval is run to compare missed recall. Search modes are semantic retrieval, full-text retrieval or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a separate switch that can be turned on. Minimum relevance takes effect only in semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. A retrieval result does not confirm the authorization status of any distributor. The citation upper limit is measured in tokens. The output citation content is an array whose length can be 0, and the downstream path still runs when nothing is retrieved. Question optimization performs coreference resolution and question expansion and adds one model call.
Hypothetical paper exercise sample: the source material reads "Complaint form CX-2024-018 records that a store in a certain region sells goods whose outer-package inkjet code does not match the dispatch batch of the authorized channel; acceptance status is pending verification", "the anti-counterfeit code record shows the inkjet code corresponds to a dispatch batch sent to another region", and "the penalty notice has not yet been issued". The draft sentence to be judged is "the store has been found to have diverted goods and the penalty has been executed". The judgment requires the same subject, the same period and the same field: the acceptance status is pending verification and the penalty notice is unsigned, so the draft sentence's claim of "found and executed" is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding diversion complaint forms and anti-counterfeit code records | Check that the vector model identifiers match before associating; compare recall differences against a base holding penalty notices |
| Search mode | Hybrid retrieval | Complaint forms contain identifiers and natural-language description; compare candidate coverage against semantic retrieval alone |
| Citation upper limit | Set to a token amount covering the complete fields of one complaint form | Test content completeness with complaint forms of different lengths; compare truncation against a smaller value |
| Minimum relevance | Take 0.3 as a hypothetical starting point | Label relevant and irrelevant samples first; compare the count of irrelevant fragments against 0.5 |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject has been altered |
| Result reranking | Enable when a rerank model is available | Compare candidate ordering before and after reranking and observe whether identifier queries improve |
Three common errors
- Writing a pending-verification status as a finding. The input is a complaint form whose acceptance status is "pending verification"; compare the draft sentence "has been found to have diverted goods"; observe whether the drafting result introduces a conclusion that is not supported.
- Treating an unsigned penalty notice as executed. The input is "the penalty notice has not yet been issued"; compare the draft sentence "the penalty has been executed"; observe whether an incomplete action is written as a completed fact.
- Inferring authorization status from a similarity score. The input is citation content after minimum-relevance filtering; compare the draft sentence's statement about the store's authorized channel; observe whether a retrieval score is used as a basis for permission or entity isolation.
How to verify the setup
- Input the three source sentences and the draft sentence from the hypothetical sample; compare the "found and executed" claim in the draft sentence against the field values "pending verification" and "not yet issued" in the materials; the condition is met when the draft sentence is observed to be unsupported and the evidence gap is recorded.
- Input a query containing an inkjet code; compare the candidate content from hybrid retrieval against semantic retrieval alone; the condition is met when the record corresponding to the code is observed in the citation content; when it does not appear, record it as pending verification.
- Input a paper empty result array; compare how the downstream node handles citation content of length 0; the condition is met when the path is observed to continue and no retrieval hit is fabricated. All of the above are paper checks and were not actually executed.