What the documents look like
Creator sales campaigns depend on external creator fulfillment and commission settlement. Brand commerce staff, MCN agencies and creator managers work from creator cooperation agreements, commission setting sheets, sample shipment forms and livestream performance reports. Procurement and supplier materials sit in the same knowledge base: supplier quotations, supply lead times, minimum order quantities and stock confirmation records. When both material groups coexist, retrieval terms can pull across them. A commission rate in a creator agreement and a unit price in a supplier quotation are different fields; if retrieval results interleave them, a reviewer may read a quotation validity period as a supply period. Materials update per session and weekly. This page treats timeliness only at that stated frequency and does not infer batch synchronization or real-time mechanisms.
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. Confirm the vector model identifier first and exclude non-matching bases; do not run cross-model retrieval to compare missed recalls. Semantic retrieval, full-text retrieval and hybrid retrieval are compared against test questions; no single mode is presumed suitable for all materials. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than an enable switch. Minimum relevance 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 same-vector-model requirement for associated knowledge bases does not extend to external systems. The reference cap is measured in tokens, and referenced content is output as an array whose length can be zero.
Image retrieval runs only when an image input and the corresponding model capability are present, and reranking runs only when a rerank model exists and is enabled by configuration. Optional paths in the overall retrieval flow are not executed on every request.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this page's sample. The numeric values are hypothetical starting points for this trial, with a stated reason and a comparison object; they are not official defaults.
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Procurement and supplier materials base | Check the vector model identifier first and exclude bases on different vector models |
| Search mode | Hybrid retrieval | Compare with pure semantic and pure full-text results, observing numbered and natural-language questions |
| Reference cap | Set to 800 tokens for this trial | Compare content completeness using supplier materials of different lengths |
| Minimum relevance | Set to 0.4 for this trial | Label relevant and irrelevant samples first, then compare against this threshold |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled per configuration | When a rerank model is available, compare rerank scores with vector scores |
Three common errors
- Treating retrieval as authorization. Retrieval returns referenced content only; business verification, approval, database writes and external synchronization each require their own downstream implementation. A pending review does not demonstrate a lack of operating permission.
- Treating minimum relevance as an entity or permission filter. A similarity score does not exclude unrelated businesses or unauthorized records.
- Treating a quotation validity period as a supply period. A contradiction requires explicitly mutually exclusive values in the same field; an unstated field or evidence covering only a related stage is unsupported or pending verification.
How to verify the setup
Hypothetical paper exercise sample. Source text: Supplier A issued a quotation on 1 March 2025 stating a unit price of 12 yuan for a given model of goods, with the quotation valid until 31 March 2025. The quotation does not state a supply start or end date, nor a minimum order quantity. Draft sentence to be judged: Supplier A commits to continuing supply at a unit price of 12 yuan from April 2025. Authorization boundary: this sample is used only to check retrieval and recall, and does not involve signing, cooperation reaching or contract effectiveness.
- Input the quotation source text and the draft sentence, compare the unit price field with the validity period field, and observe whether the draft sentence treats the end of validity as a supply commitment. The draft sentence exceeds what the source states and is unsupported.
- Input "Supplier A unit price" as the retrieval term and compare whether the returned referenced content includes the quotation source text. Returned content length can be zero; when an empty result appears, record the gap and do not assert that a fabricated rule necessarily returns an empty array.
- Input "supply start and end date" as the retrieval term and compare whether the returned referenced content states that field. The source does not state it; record this as pending verification and do not fabricate authorization or effectiveness clauses to obtain a definite answer.
For each check, the material needed for verification must lie within the target retrieval scope to which access is held, or within an explicit manual input; when it is absent, record the gap. Retrieval checks describe observation and failure branches and do not guarantee a first hit, full recall, or an empty result for an unrelated query. Empty-result handling can use an explicit paper input of [] to check downstream design, without claiming that a fabricated rule necessarily returns [].