What the documents look like
Trade returns and claims handle reverse logistics and liability determination after export. The working materials include return agreements, claim letters, quality reports and return customs declarations. These arrive from customer email, inspection reports, customs systems and insurance claim systems. Each case assembles a different combination of documents, and updates follow case occurrence rather than a fixed schedule.
For copy drafting, these materials set the ceiling on what a product or marketing sentence can state. The return reason, the liability assignment, the compensation status and the customs progress each sit in a different document. A supporting phrase in draft copy must trace back to a specific passage. Facts absent from the materials cannot be filled in during drafting. The retrieval step brings back the passages connected to the sentence under draft so the drafter can compare them.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. At processing time, check the vector model identifier first, exclude non-matching bases, and record the observed result. Do not run a cross-model search to compare missed recall.
Retrieval output is an array of references, and its length can be 0. When nothing is found, the output path still completes. Semantic retrieval, full-text retrieval and hybrid retrieval are compared separately against test questions. No single mode is presumed suitable for all materials. Numbers, models and proper nouns suit full-text or hybrid retrieval; natural-language phrasing suits semantic or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch that can be turned on. Reranking requires an available model and must be enabled according to settings. Image retrieval requires corresponding image input and model capability. Optional paths shown in the overall document flow diagram do not mean every path executes on every request.
The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation. Retrieval scope is determined by the associated knowledge bases, and a similarity score does not distinguish business ownership.
Hypothetical paper exercise sample (constructed on paper only, involving no real customer or identifier):
Material passage one: "Return agreement number TR-2024-0117 states that the buyer raised an appearance nonconformity after signing on 12 March 2024, the two parties agreed to return 200 units, and the return reason field records transport damage."
Material passage two: "The conclusion paragraph of the accompanying quality report states that among 40 sampled units, 6 showed carton compression marks, and all internal units passed functional testing."
Draft sentence to be assessed: "This batch was returned due to a quality defect, and the party making this statement has confirmed a compensation plan."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the bases holding return agreements, quality reports and claim letters | Verify these bases use the same vector model; bases on other vector models are not associated together |
| Search mode | Start with hybrid retrieval | The draft sentence contains natural-language wording such as "quality defect" while the materials contain numbers; compare candidates from both paths |
| Reference limit | Set a trial starting point of 1500 tokens | Return agreements and quality report passages are long; test with materials of different lengths to see whether content stays complete, then compare against a smaller value |
| Minimum relevance | Set a trial starting point of 0.4 | Label relevant and irrelevant samples first, observe whether irrelevant fragments are filtered at this value, then compare against adjacent values |
| Question optimization | Leave off in the first round; compare on and off during multi-turn follow-ups | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Compare ordering after enabling according to settings | When candidates are numerous, observe the difference between rerank scores and vector scores |
Three common errors
- Writing a conclusion into the draft sentence that the materials do not state. The passages above record only transport damage and appearance compression marks, with functional testing fully passed. A draft sentence that says "quality defect" and asserts a confirmed compensation plan is unsupported. During review, compare each sentence against the subject, object, status and effective scope recorded in the source text.
- Treating a similarity setting as a permission decision. The minimum relevance value filters only low-relevance results; it does not exclude unrelated business or unauthorized records. A retrieved passage shows only that it was recalled, not that the drafter has the right to use it.
- Blurring pending review, quotation validity and signature status into one. Pending review does not demonstrate a lack of operating permission; quotation validity does not indicate a supply period; signing, reaching cooperation and contract effectiveness are three distinct facts. Evidence gaps stay as gaps.
How to verify the setup
- Input the material passages above together with the draft sentence, and compare the two expressions "quality defect" and "confirmed a compensation plan" against what the source text records. The source text records only transport damage and appearance compression marks, and no compensation confirmation record. The condition is not met when such unsourced expressions appear.
- Input identifier TR-2024-0117 from material passage one, and compare recall results under full-text retrieval and semantic retrieval separately. Observe whether the identifier is matched, and whether semantic retrieval returns passages that are semantically close but carry a different identifier. When the identifier is not matched, record that branch; full recall is not guaranteed.
- Use an explicit paper-input empty array to check the downstream design, observing whether the path completes when reference length is 0 and how downstream stages receive it. Do not claim that a fictional rule search necessarily returns an empty array. To verify whether a compensation plan exists, a corresponding document must be found within an accessible target retrieval scope or an explicit manual input; when absent, record it as an evidence gap.