What the documents look like
Platform refund and return handling happens order by order, not as a single concentrated event tied to one live-stream commitment. The working materials come from the after-sales backend, customer service chat, WMS receiving, and payment refund records. For procurement and supplier files, the daily documents are refund applications, return logistics notes, quality inspection photos, and refund approvals. These records update per order and per day, and the refund application, inspection photo, and approval result for one order may sit at different points in time.
For the vector model and index step, whether a record can be retrieved depends on whether training and index creation were completed at ingestion. Uploading an image at query time does not show that an image index already exists in the library. One source data block can map to several indexes, and the index content and the original text need separate checks. Linking retrieval libraries requires the same vector model, and that requirement applies to the knowledge base link only, not to outside systems.
Constraints at the vector indexes step
The citation limit is measured in tokens, not in a top-k count. The minimum relevance value is a filter threshold that applies only in semantic retrieval or when result reranking is used; it does not act as a permission or entity isolation control. Hybrid retrieval fuses multiple result paths through RRF, which is a fusion process rather than a separate toggle. Image retrieval needs an image input at query time and a model with the matching capability; when no image description or image vector was generated at ingestion, a pure image query is largely unusable. Reranking needs a rerank model enabled in settings, and that path does not run when it is not enabled.
When retrieval returns nothing, the node still outputs a citation array whose length can be zero, and the downstream path continues. An empty result does not mean the corresponding business record does not exist.
Configuration and implementation checks
The table below uses one hypothetical drill sample for this page and lists only documented settings or implementation checks tied to this feature and the procurement task. The numbers are starting points for this drill, with the reason for choosing them and the comparison target.
| Configuration item | Hypothetical value for this drill | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the procurement and supplier library that uses the same vector model | Check the vector model identifier of each library first, exclude those with a different identifier, and do not run cross-model retrieval comparisons |
| Search mode | Hybrid retrieval | Supplier names and purchase order numbers coexist with natural language questions, compared against a single semantic retrieval result |
| Minimum relevance | Hypothetical starting point 0.35 | This value is a starting point for this drill, not an official default; compare cited original text item by item against 0.2 and 0.5 before adjusting |
| Citation limit | Hypothetical starting point 2000 tokens | This value is a starting point for this drill; compare citation completeness against 1000 and 3000 tokens |
| Question optimization | Enabled for multi-turn follow-up scenarios | Observe a single clear question and a multi-turn follow-up separately, and compare cited original text before and after retrieval term completion |
| Image index check | Check whether an image description or image vector was generated at ingestion | No numeric value; check whether the ingestion record holds the corresponding index, judged separately from whether an image was uploaded at query time |
Three common errors
- Treating minimum relevance as a permission or entity isolation measure. The value is a filter threshold and does not distinguish supplier ownership or authorization scope; passing retrieval does not mean access has been authorized.
- Linking libraries with different vector models and then inferring runtime behavior. The documentation only requires that linked libraries use the same vector model and does not describe recall behavior after linking libraries with different models, so such an inference lacks support.
- Treating an image uploaded at query time as proof that an image index already exists at ingestion. Whether an image can be retrieved depends on whether an image description or image vector was generated at ingestion, and the two stages need separate checks.
How to verify the setup
The following uses the same hypothetical drill sample. The source material reads: "Purchase order CG-2024-0912, supplier East China Packaging Materials Co., Ltd., purchased item five-layer corrugated cartons, quantity 800, ingestion date September 12, 2024, inspection conclusion qualified. Refund application R-2024-1031 corresponds to order O-2024-7788, returned item 120 cartons from the same batch, return logistics note signed for, refund approval status pending review." The draft sentence to be judged reads: "The returned item for this order has completed refund, and the supplier's supply responsibility has been confirmed."
- Input the source material and the draft sentence, and compare whether "refund approval status pending review" and "has completed refund" point to the same order and the same field. A contradiction exists only when the approval status field shows a clearly exclusive value; under the current evidence the sentence is unsupported and remains to be verified.
- Input the source material and "the supplier's supply responsibility has been confirmed," and compare whether the source states a responsibility conclusion or an authorization basis. The source records only a qualified inspection and the returned quantity, with no responsibility attribution, so the sentence lacks evidential support.
- Input the same question group against retrieval with question optimization on and off and against different minimum relevance values, and compare whether the returned cited original text contains the purchase order and refund application above. Observe whether citations are empty or unrelated to the question, record the actual result, and do not presuppose that the first item always matches or that an unrelated query always returns empty.