What the documents look like
Trade returns and claims work sits after export, in the reverse logistics and liability stage. The materials come from customer email, quality inspection reports, customs systems and insurer claim systems. The working documents are return agreements, claim letters, quality reports and return customs declarations. These are maintained by after-sales, customs affairs and insurance claim roles, and they are updated irregularly, once per case. The formats differ sharply. A return agreement carries parties and liability terms. A claim letter carries a requested amount and its basis. A quality report carries batch inspection conclusions. A return customs declaration carries a declaration number and cargo details.
Product data and selection questions must be answered by retrieving across these four document types. Retrieval output is reference content only. It does not constitute authorization, approval or a claim decision.
A hypothetical paper exercise sample is used below for illustration only. A trade after-sales staff member submits the original text of a return agreement to check whether selection question answers cite it accurately. The original text reads: "Party A is a domestic exporting enterprise, Party B is an overseas buyer. The parties agree that batches returned for quality non-conformity are governed by the conclusions of the quality report, and that Party A handles the return declaration. This agreement takes effect from the date both parties sign." A draft sentence to be assessed reads: "The return agreement has been signed and has taken effect by both parties, and the return declaration has been completed by Party A." The draft sentence covers three facts: signing, taking effect and completion of the declaration. The original text states only the agreement and the condition for taking effect. It does not record a signing fact or a declaration completion record.
Constraints at the retrieval step
The knowledge base search node can be associated with one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Check the vector model identifier first and exclude those that do not match.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Numbers, model codes and proper nouns suit full-text or hybrid retrieval. Natural language questions suit semantic or hybrid retrieval. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. After hybrid retrieval, reranking is usually needed and filtering is done by the rerank score. Question optimization performs coreference resolution and question expansion, and it adds an extra model call.
Image retrieval requires an image input and a model that supports it. Reranking requires a model and must be enabled by configuration. These optional paths do not mean every request executes all of them. When retrieval returns nothing, the reference content is output as an array and its length can be 0, and the downstream chain still runs. Retrieval and similarity scores do not carry permission or entity isolation functions.
Configuration and implementation checks
The table below covers the sample above and lists documented settings or implementation checks relevant to this feature and the primary task. The values are hypothetical starting points for this exercise, used only for comparison. They are not official default values.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the trade return and claim base that uses the same vector model | Check each base's vector model identifier first, exclude non-matching ones, then associate |
| Search mode | Run one round each for full-text retrieval and hybrid retrieval | Use the same question and observe the difference in hits for number-type and natural language questions |
| Reference limit | Set 2000 tokens for now | Use one passage from each of the four document types with different lengths as input, and compare whether reference content is truncated |
| Minimum relevance | Set 0.4 for now | Label relevant and irrelevant samples first, then observe whether irrelevant content is filtered at this value |
| Question optimization | Off in the first round, on in the second | Use multi-turn questions containing references, and compare whether the retrieval terms change the subject before and after optimization |
| Result reranking | Not enabled in the first round, enabled in the second | When candidates are numerous, compare the rerank score with the original similarity ordering |
Three common errors
- Treating a retrieval hit as authorization or completed execution. The draft sentence above turns an agreement term into "has been signed and has taken effect, and the declaration has been completed." Retrieving the agreement original text shows only that the material exists. Signing and declaration completion are separate facts and require corresponding records. Check: input the agreement original text and the draft sentence, compare the three facts in the draft sentence against the scope stated in the original text, and the draft sentence fails if it states a completion status not recorded in the original text.
- Treating minimum relevance as a permission or entity filter. This value takes effect only under semantic retrieval or result reranking, and it filters low-relevance results. It does not exclude unrelated businesses or unauthorized records. Check: input a question unrelated to trade returns, compare the returned reference content, and it fails if an empty result is interpreted as permission isolation.
- Associating knowledge bases with different vector models together. The documentation requires associating knowledge bases that use the same vector model. Check: input two knowledge bases with different vector model identifiers, compare whether the identifiers match, and association still being performed when they do not match fails.
How to verify the setup
- Use the return agreement original text and the draft sentence above as input. Compare the three facts in the draft sentence against the scope stated in the original text, and observe whether the system reference presents only the original text content. If the reference content supports "has been signed and has taken effect" while the original text records no signing fact, the result is judged unsupported.
- Use a question of the "return customs declaration number" type as input. Compare the reference content from full-text retrieval and hybrid retrieval, and observe whether the number is matched completely. If the reference content omits the number or returns only a general description, record the observation and check chunking and retrieval terms.
- Use a multi-turn question containing a reference as input. Compare the retrieval terms with question optimization off and on, and observe whether the subject is changed incorrectly. If the optimized retrieval term replaces the original subject, record the observation and adjust the question optimization setting.
An empty result handoff can be checked against the explicit paper input [] to verify downstream design. It cannot be claimed that a given rule always returns [] on retrieval. The same-vector-model requirement for retrieval bases applies only to knowledge base association and does not extend to business registration sites or other external systems. A pending review does not demonstrate a lack of operating permission. Signing, reaching agreement and contract effectiveness are three distinct facts.