What the documents look like
Semi-managed sellers on Temu maintain overseas inventory and ship from overseas warehouses, so operational records exist on two sides at once: platform-side declarations and overseas-warehouse physical records. Four material types matter for this task: product declaration forms, overseas inventory tables, price declarations, and platform fulfillment orders. Two update rhythms coexist, daily and per order, and the same transaction leaves one record at each of the four stages of declaration, inventory, price, and fulfillment. Contract and transaction material comparison examines whether subject, object, status, and validity scope agree across these four material types. At the retrieval step the input is the text of these materials and the output is cited content, whose length can be zero. Retrieval itself does not change the authorization status of any material.
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. When candidate bases are presented, check the vector model identifier first, exclude bases whose identifier does not match, and record the association result actually observed.
Semantic retrieval depends on vector distance; full-text retrieval depends on keyword hits; hybrid retrieval uses both paths and merges them through RRF. RRF is a fusion process, not a toggle. 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. The citation cap is measured in tokens, so materials of different lengths are truncated to different degrees under the same cap. Question optimization performs coreference resolution and question expansion and adds one model call before formal retrieval.
Hypothetical drill sample: on 10 March 2026, a semi-managed seller operations staff member prepares a handover request. The source material text consists of three sentences. First: "The overseas inventory table (1 March 2026 to 7 March 2026) records SKU-A as having 120 units available in the Germany warehouse." Second: "The platform fulfillment order (generated 5 March 2026) records SKU-A as having 40 units pending shipment from the Germany warehouse." Third: "The price declaration (effective 1 March 2026) records SKU-A at a declared price of 19.99 EUR on the Germany site, valid through 31 March 2026." The draft sentence to be judged is: "SKU-A has 120 units shippable from the Germany warehouse, a declared price of 19.99 EUR, and a supply term covering the whole of March 2026." Checkpoints: the inventory table and the fulfillment order differ in period and field, so 120 units and 40 units are not mutually exclusive values in the same field; the price declaration states only a quotation validity period and does not state a supply term, so the supply term is unsupported.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding product declaration forms, overseas inventory tables, price declarations, and platform fulfillment orders | Check each base's vector model identifier first, exclude non-matching bases, and do not run cross-model retrieval comparisons |
| Search mode | Run semantic retrieval once, then full-text retrieval once | Compare the cited content of both paths on the same test question; observe full-text hits on identifier-type fields |
| Citation cap | Starting point set to 2000 tokens | This is a hypothetical starting point for this trial; compare materials of different lengths to see whether content is truncated, then decide the adjustment direction |
| Minimum relevance | Starting point set to 0.5 | Manually label relevant and irrelevant samples first, then compare cited content before and after filtering on the same question |
| Question optimization | Enabled only for multi-turn follow-ups or short questions | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enabled when candidate results are numerous | Compare ordering for the same question before and after reranking; the rerank score is used for filtering |
Three common errors
- Reading the 120 units in the inventory table and the 40 units in the fulfillment order as a contradiction. The two differ in field and period; only mutually exclusive values in the same field constitute a contradiction, and this case is unsupported.
- Reading the price declaration's validity date of 31 March 2026 as a supply term. A quotation validity period does not indicate a supply term, so the draft sentence's "covering the whole of March 2026" lacks material support.
- Treating a retrieval hit as authorization. Retrieval returns cited content and does not indicate that the handover request has been approved; a pending-review status does not demonstrate a lack of operating permission, and signing, reaching cooperation, and contract effectiveness are three distinct facts.
How to verify the setup
- Enter the three source sentences above and the draft sentence, and compare the quantity, price, and term in the draft sentence against the corresponding fields in the source. If the draft sentence writes 120 units as shippable quantity and writes the quotation validity period as a supply term, the sentence is unsupported and the evidence gap should be recorded.
- Enter the same test question and run semantic retrieval and full-text retrieval separately, then compare the cited content of both paths. If an identifier-type field is not recalled under semantic retrieval and is recalled under full-text retrieval, record that observation; a first hit or complete recall is not guaranteed.
- Enter an explicit paper empty result
[]and check how the downstream chain is designed to accept an empty citation. The empty-result handoff can be checked with this input; do not claim that a fictional rule retrieval necessarily returns[].