What the documents look like
Platform product title and detail compliance work concerns the truthfulness of promotional wording and the display of qualification material. It does not concern parcel delivery, which belongs to platform shipping operations. The working documents in this category are product title tables, detail page assets, advertising prohibited-word lists, and quality inspection reports. These four document types arrive in different shapes: title tables are row-based and short-field, detail page assets mix images with caption text, prohibited-word lists are numbered clauses, and inspection reports carry test items, conclusions, and validity periods.
Updates occur weekly and at each new product launch. A single product's title, assets, prohibited-word list, and inspection report can therefore sit at different versions at the same moment. Document parsing and chunking determines whether these materials can be recalled accurately once they enter the knowledge base. If chunking separates a clause's condition from its conclusion, or splits a scope statement away from the term it constrains, a selection question will quote an incomplete statement. Retrieval results reflect document content only. They do not constitute compliance approval, nor do they authorize a delisting or takedown.
Constraints at the document chunking step
Materials mix tables, clause lists, and reports, so chunking must keep the clause subject, the applicable product scope, and the effective period inside the same data block. In the documented storage structure, one group of original data can map to multiple vector indexes. Original data and retrieval indexes live in separate stores, and a successful upload is a different event from an index being searchable. The smallest search unit is the library, not the collection; collections classify data and do not affect search results.
Hypothetical paper exercise sample. Source text for comparison: "Advertising Prohibited-Word List, March 2026 edition, Clause 12: Ordinary food detail pages must not contain the expressions 'treat' or 'cure'. Applicable scope: snack category. Effective date: 1 March 2026." The draft sentence to be assessed: "This nut product's detail page uses the word 'relieve' to describe the eating experience." The assessment compares the wording of the draft sentence against the expressions listed in Clause 12 and against the clause's applicable scope. The sample does not state whether "relieve" appears on the prohibited-word list, so that conclusion remains pending verification. A separate handover request reads: "Please submit the above draft sentence to legal review." That request is in a pending-review state and does not constitute a review pass.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | Clause lists contain numbers and proper nouns, so semantic recall and full-text recall run together; compare against a single-path result |
| Minimum relevance | 0.3 | Hypothetical starting value for this exercise, used to observe candidate count after filtering; compare against running with no threshold |
| Reference limit | 800 tokens | Hypothetical starting value, estimated from one clause plus its explanatory text; compare against a smaller value for reference completeness |
| Question optimization | Enabled | The draft sentence contains a reference; observe whether the completed query covers the product category |
| ReRank | Enabled | Second-pass ordering when candidates are numerous; compare clause position before and after reordering |
| Chunk review | Manual review | Check whether clause subject, applicable scope, and effective date stay in the same data block |
The reference limit is measured in tokens, not in a count of items. Minimum relevance is a filter threshold and applies only under semantic search or when result reordering is used; it is not a final ordering rule, and it does not exclude other categories or act as an authorization check. ReRank produces a 0-1 score, and RRF is the fusion step that combines multiple recall paths; it is not a switch that is turned on. Image retrieval requires an image input and matching model capability, and reordering requires an available model enabled by configuration. Optional paths in the overall retrieval flow do not all execute on every request.
Three common errors
- Treating a successful upload as an updated index. Upload, training, and index creation are separate stages. Before retrieval, check the actual training state and whether an index was generated.
- Treating a similarity score as a category or permission filter. Minimum relevance only filters low-relevance results. It does not exclude clauses from other categories and does not perform authorization judgment.
- Treating a pending-review request as approved. While the handover request sits in pending review, the compliance conclusion for the draft sentence remains undetermined, and no review conclusion record exists.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence, then compare whether the retrieved clause list includes Clause 12's subject and its applicable scope. If the returned content is missing the applicable scope or the effective date, chunking has separated the clause condition, and the data block needs rechecking.
- Input "snack category treat cure" as an exact-term query, then compare candidates from full-text search against candidates from hybrid search. If the clause does not appear among candidates, check whether that clause completed training and index creation, and record the actual observation without presuming a failure cause.
- Input an empty reference list as a paper check, and inspect how the downstream question-answering design handles output when no reference content is present. This step checks the empty-result handling logic. It does not assert that any retrieval against the hypothetical rules will return an empty list.
Documents from libraries built on different vector models cannot be associated with each other. Verify identifiers first, exclude those that do not match, and record what is actually observed. Where the material needed for a compliance judgment is absent from the accessible retrieval scope or from a clearly defined manual input, record the gap and leave the conclusion pending.