What the documents look like
Title and detail compliance work covers the truthfulness of promotional wording and the display of qualification materials. It does not cover parcel delivery, so the documents involved differ from those used in shipping operations.
The four recurring document families are the platform rule centre, the advertising prohibited-terms list, the brand asset library, and third-party inspection reports. Two working documents sit alongside them: the product title table and the detail page assets. Updates arrive weekly and with each new listing, so a handover may involve several versions of the same document.
These materials differ sharply in shape. A prohibited-terms list is mostly short entries. An inspection report carries report numbers, sample models, tested items, and a report status. A title table carries parameter wording and any absolute phrasing that needs review. Retrieval therefore has to serve two query types at once: natural-language description and exact identifiers such as a report number or a model code. Chunk lengths vary widely, with short entries and long report paragraphs in the same working set.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. Before a handover run, compare the vector model identifier of each candidate knowledge base, exclude any that do not match, and record what was observed. Do not run a cross-model search in order to compare missed recalls.
Cited content is returned as an array, and its length may be zero. The output path still runs when nothing is found, so an empty citation list is a normal result rather than a failure of the node.
The minimum relevance value is a filter threshold, and it applies only under semantic retrieval or when result reranking is used. It is not a permission filter and does not isolate entities or business units. A retrieval hit is not an authorization. Finding a prohibited-terms entry or an inspection report does not mean the operator holds the right to edit a title or replace an asset.
Hybrid retrieval merges several recall paths through RRF. RRF is the fusion step inside hybrid retrieval, not a switch that is turned on. Image retrieval requires image input and a model capable of the corresponding processing. Reranking requires a rerank model and must be enabled in the settings. The optional paths shown in the overall retrieval diagram do not all execute on every request.
Hypothetical paper exercise sample. A platform store operations role is preparing to hand over a batch of product compliance materials. The materials supplied for checking read as follows. First, version 12 of the advertising prohibited-terms list states: "Titles in this category must not use absolute expressions beginning with the character 'most'." Second, inspection report Q-2024-031 states: "Sample model A2, tested item formaldehyde release, report status issued." One handover request awaits a determination: "Please confirm that this store's title compliance operation rights have transferred to the receiving person with this handover." The request names no authorization document, no handover effective date, and no scope of rights. It remains a matter to be verified.
Configuration and implementation checks
The table below relates to the sample above. Values are starting points for this paper exercise, with the reason for choosing them and the comparison object stated. They are not official defaults.
| Configuration item | Hypothetical starting point | How to check |
|---|---|---|
| Linked knowledge bases | Prohibited-terms base and inspection report base | Compare the vector model identifiers of the two bases; exclude any mismatch and do not link them together |
| Search mode | Hybrid retrieval | Use full-text retrieval for the phrase "beginning with 'most'" and semantic retrieval for "A2 formaldehyde" |
| Citation limit | 2000 tokens as a starting point | Chosen to hold a full inspection report paragraph; compare against a 1000-token setting for content completeness |
| Minimum relevance | 0.5 as a starting point | Label relevant and irrelevant samples first, then adjust; applies only under semantic retrieval or reranking |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled per settings | Confirm a rerank model is available; the rerank score runs from 0 to 1 and can be used for filtering |
Three common errors
- Treating a retrieval hit as a basis for permission. The input is the retrieval result from the prohibited-terms base. The comparison object is the permission wording in the handover request. The observable result is that the retrieval returns term entries, while no authorization document and no effective date appear in the materials. The permission question therefore remains unverified.
- Linking knowledge bases across different vector models. The input is the vector model identifier of the prohibited-terms base and of the inspection report base. The comparison object is whether the two identifiers match. The observable result is that they differ. The correct handling is to exclude the mismatched base and record the observation, without inferring missed recalls or redundant results.
- Treating a pending status as already effective. The input is the status field of inspection report Q-2024-031. The comparison object is the value "issued" against the value "review passed". The observable result is that the source text states only "issued". This difference is unsupported by the materials and is not described as a denial or a conflict.
How to verify the setup
- Using the source wording "Titles in this category must not use absolute expressions beginning with the character 'most'", enter the search term "beginning with 'most'". Compare whether the returned entry matches the source wording. The condition is met when the returned content contains that expression and is not truncated. This is the expectation; no live test is claimed.
- Using the source wording "Sample model A2, tested item formaldehyde release, report status issued", enter the search term "A2 formaldehyde release". Compare whether the returned passage contains the model and the tested item. The condition is met when both appear. This is the expectation; no live test is claimed.
- Using the handover request "Please confirm that this store's title compliance operation rights have transferred to the receiving person with this handover", enter the request text. Compare whether an authorization document, an effective date, and a scope of rights appear anywhere in the materials. The observable result is that none of the three appears, and this is recorded as an evidence gap. A paper input of an empty array can be used to check the downstream design; no claim is made that a search of a fictional rule would return an empty array.