What the documents look like
Creator sales campaigns for procurement and supplier records centre on four document types: creator cooperation agreements, commission setting sheets, sample shipment sheets and livestream performance reports. The operating parties are brand business staff, MCN agencies and creator managers. Fulfilment and settlement depend on external creators rather than internal brand staff, so the procurement record has to hold the creator side of the deal as well as the supplier side. Materials update per session and per week. The same supplier's quotation, sample dispatch and commission terms may sit in different documents, and chunking has to preserve document identity and version so that a condition, its explanation and the content it governs are not split apart. The smallest search unit is the library; a collection only classifies data and does not affect search results.
A sample shipment sheet carries a shipment number, a shipment date and a quantity. A commission setting sheet carries a rate and a validity period. Neither document type, as described here, carries a signed receipt field or a settlement status field. That gap matters at chunking time, because a chunk that merges a shipment sheet with a commission sheet can make a draft sentence look supported when the underlying fields are absent.
Constraints at the document chunking step
Numbers, models and proper nouns in these materials suit full-text search or hybrid search. Natural-language supply conditions suit semantic search together with question optimisation. One set of original data can map to multiple indexes, and the completeness of a segment directly affects recall quality. Whether an image-based sample sheet or performance screenshot can be retrieved depends on whether an image description index or image vector index was built at ingestion, and on whether the vector model accepts image input. Hybrid search merges multiple result paths through RRF; reranking needs a model and is enabled according to settings. Retrieval does not perform authorisation checks, and a similarity score does not distinguish suppliers or cooperation status.
A hypothetical paper exercise sample: the material text for checking reads "Party A: Xinghe Trading (sample entity); Party B: Creator A Studio. Sample shipment sheet number SP-2024-0001, shipment date 2024-05-10, quantity 2 items. The commission setting sheet states a commission rate of 15% for this session, valid until 2024-06-30." The draft sentence to be judged reads "Xinghe Trading and Creator A Studio have completed shipment of sample shipment sheet SP-2024-0001 and confirmed commission settlement." The source text states only a shipment date and a quantity. It states no signed receipt and no settled commission.
Configuration and implementation checks
| Check item | Relevant setting or implementation check for this page | Hypothetical trial starting value and comparison method |
|---|---|---|
| Search mode | Semantic search, full-text search or hybrid search | Search the number SP-2024-0001 by full-text search and compare hits against a natural-language question |
| Reference limit | Each search references at most n tokens of content | Starting value 3000 tokens, compared with 1500 tokens, to see whether referenced content covers commission conditions |
| Minimum relevance | A 0-1 value, effective only in semantic search or with reranking | Starting value 0.3, compared with 0.5, to compare result count and relevant fragments |
| Question optimisation | Coreference removal and question expansion before retrieval | Enable and disable for "what is the commission for this session" and observe whether the subject is completed |
| Result reranking | ReRank model performs a second ordering, score 0-1 | Enable when candidates are numerous; compare ranking position before and after reranking |
| Post-update review | After a material change, check actual training and index state | After uploading a new version, check whether retrieval returns the new content; do not treat a successful upload as an updated index |
The 3000-token and 0.3 starting values are hypothetical trial values chosen for this page's comparison, not documented defaults. The comparison method is a paired observation of referenced content and result count at two values.
Three common errors
- Treating the shipment date and quantity on a sample shipment sheet as a signed receipt. The source text states only shipment date 2024-05-10 and quantity 2 items, with no receipt field. The phrase "completed shipment" in the draft sentence lacks evidence.
- Treating commission rate 15% and validity until 2024-06-30 as a settled commission. The commission setting sheet states a rate and a validity period, with no settlement status. The phrase "confirmed commission settlement" is unsupported.
- Treating a retrieval hit as an authorisation pass. Returning a cooperation document does not mean the supplier is approved or the contract has taken effect. Signing, reaching cooperation and contract effectiveness are three distinct facts.
Corresponding checks:
- Input the draft sentence and the sample shipment sheet text; compare whether a receipt field exists. If the source has no receipt record, the claim is unsupported.
- Input the draft sentence and the commission setting sheet text; compare whether a settlement status is stated. If only a rate and validity period are stated, record an evidence gap.
- Input the retrieval request and the returned documents; compare whether an authorisation field is present. If only document content is returned, record that retrieval is not authorisation.
How to verify the setup
- Input the shipment date and quantity of SP-2024-0001 from the hypothetical source text; compare whether the draft sentence exceeds the source scope. If the draft claims a signed receipt while the source has no such field, the claim is unsupported.
- Input the commission setting sheet text; compare whether the draft sentence equates validity until 2024-06-30 with completed settlement. If the source states no settlement status, the item remains pending verification.
- Input the paper input [] for empty-result handling; check whether the downstream design treats it as an empty reference. Do not claim that a hypothetical rule's retrieval necessarily returns []. Cross-model association is not available; check identifiers first, exclude non-matching items, and record the actual observation.