What the documents look like
Cross-border product compliance work happens before listing and before import. The working materials are product test reports, English labels, MSDS files and platform prohibited-item lists. These arrive from platform policy centers, testing laboratories, destination-country regulation repositories and supplier submissions. Updates are irregular, and each new product listing can produce new versions of the same material.
For operations handover, the effective scope, issuing party and version status of each material determine what may be cited and what may be used to move a step forward. A test report signed by a laboratory supports the test facts it covers. An English label in draft status does not carry the same standing as a signed label. An MSDS and a platform prohibited-item list each answer different questions. The citation tracing step exists to keep these distinctions visible when responsibility passes between operators.
Once materials enter retrieval, the knowledge base search node outputs cited content as an array, and the length may be zero. An empty array means the output path still runs; it does not mean the workflow has errored or terminated. The citation limit controls the maximum amount of content cited per search, measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. These parameters affect the quantity and ordering of cited content. They have no connection to authorization status. Retrieval results pass to the answering step, where each answer is checked against the cited source text. The source evidence is the cited original text.
Constraints at the citation tracing step
The following is a hypothetical drill sample, used only to illustrate a checking method. It does not point to any real entity or record.
Material text available for checking: A supplier submitted an English label draft on 1 March 2026, marked as applying to model A-100, with issuance status draft, and without a testing laboratory signature page. The same supplier submitted a test report for that model on 5 March 2026, with status signed by the testing laboratory, covering the ingredient items stated on the label.
A handover request to be judged: Use the 1 March English label draft as the label basis for citing A-100 at listing, and mark it in the handover record as effective.
Judgment conditions: The label draft and the test report belong to different fields. Draft status and signed status are mutually exclusive values of the same field. The request treats the draft as an effective label basis, which does not match the issuance status recorded in the material text. This is an unsupported citation. The signed test report supports only the ingredient-item test facts. It does not show that the label itself has taken effect. The material text does not state a label effective date or authorization terms. That part remains an evidence gap.
Retrieval and authorization are separate matters. Knowledge base search may return cited content for the label draft. That does not confirm the label is effective, and it does not authorize a listing action. A similarity score and the minimum relevance value do not perform entity isolation or permission filtering. No conclusion can be drawn from them that a type of entity lacks access or that a record has been excluded. Linking knowledge bases requires the selected knowledge bases to use the same vector model. That condition constrains only which knowledge bases may be linked together. It does not extend to external regulation repositories or business registration systems also needing the same model. Hybrid retrieval merges multiple result paths through RRF. This is a merging process, not a switch that can be turned on. Image retrieval requires image input to exist and the model to have the corresponding capability. Reranking requires a model and must be enabled by configuration. The paths shown in the documentation flow diagram do not mean every request executes all of them.
Configuration and implementation checks
| Configuration or check item | Setting in this page's sample | Basis and comparison method |
|---|---|---|
| Linked knowledge bases | Select only knowledge bases sharing the same vector model as the compliance materials | Official documentation states only knowledge bases with the same vector model may be linked; verify the identifier first and exclude those that do not match |
| Citation limit | Starting value for this hypothetical drill set at 3000 tokens | This value is a drill starting point, not an official default; compare against the total length of the sample material passages and adjust |
| Minimum relevance | Starting value for this hypothetical drill set at 0.3 | This value is a drill starting point, not an official default; it takes effect only under semantic retrieval or result reranking and is unrelated to the validity of cited original text |
| Citation output handling | Pass the search node output to the answering step and check each answer against the cited original text | Official documentation states the output is an array and its length may be zero; an empty array still runs downstream |
| Material version check | Record issuance status, applicable model and submission date for each material | Manual check action; in the sample, draft and signed are mutually exclusive values of the same field |
| Authorization status record | Distinguish pending review, signed and effective in the handover record | Manual check action; pending review does not show a lack of operating permission, and signing and taking effect are different facts |
Three common errors
- Treating a label draft as an effective basis. The input is the 1 March English label draft. The comparison object is the issuance status field of the same material, where draft and effective are mutually exclusive values. The observed result is that the two do not match, and the citation request is unsupported.
- Inferring that the label has taken effect from a signed test report. The input is the 5 March test report. The comparison object is the report's coverage against the ingredient items stated on the label. The observed result is that the report supports only ingredient-item test facts, while the label effective status is not stated in the material and remains an evidence gap.
- Treating a retrieval hit as an authorization action. The input is the citation array returned by the search node. The comparison object is the authorization status field in the handover record. The observed result is that the presence of cited content does not change the authorization status, which remains pending verification.
How to verify the setup
- Input the 1 March English label draft text and the handover request to be judged. Compare the values of the issuance status field. The observed result is that the request's marked effective status does not match the draft status in the original text, and the request is judged unsupported.
- Input the 5 March test report text and the ingredient items stated on the label. Compare the report's coverage. The observed result is that the report supports ingredient-item testing, the label effective status is not stated, and this is recorded as an evidence gap.
- Input the citation array output by the search node. Compare the authorization status field in the handover record. The observed result is that cited content is present while the authorization status is unchanged, so authorization remains pending verification. If the citation array is empty, check whether the downstream design continues to run on an empty array, and do not infer that retrieval will necessarily return an empty array.