What the documents look like
TikTok Shop US store operations sit under local return windows, return grading and overseas logistics constraints. The working materials are product listings, return and refund records, return grading evidence, and settlement statements. These materials update daily and per order. A single record often mixes Chinese operations notes with English platform fields, so a question may be asked in Chinese while the supporting passage is written in English. Semantic retrieval in FastGPT is documented as capable of cross-language understanding, for example matching a Chinese question against an English knowledge point. That capability matters directly for reconciling multilingual materials in this category.
The 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. Collections inside a library are for classification and management; the smallest search unit is the library, so a search runs across the whole library.
Constraints at the retrieval step
One knowledge base search is not a single path from question to vector store to result. FastGPT combines question optimization, semantic retrieval, full-text retrieval or hybrid retrieval, image retrieval, RRF fusion, and Rerank according to the input content and search parameters, then applies similarity filtering and a reference cap before returning reference content. The output is an array and its length may be zero; the output link still runs even when nothing is found.
Hybrid retrieval produces a large candidate range and cannot be directly similarity-filtered, so a rerank pass is usually needed and the rerank score is used for filtering. The minimum relevance value only takes effect under semantic retrieval or when result reranking is used. It is a filtering threshold and does not carry permission or entity isolation. RRF is a fusion step over multiple result channels, not a switch. Image retrieval requires an actual image input and the corresponding model capability; reranking requires a rerank model and its configured enablement. The optional paths in the overall flow are not all executed on every request.
Hypothetical paper exercise sample. One record in a merchant knowledge base reads: "The return request for Product A was submitted on 4 March 2026, status pending review; the return window closes on 10 March 2026." Another record in the same library reads: "The return request for Product A was submitted on 4 March 2026, status approved." The draft sentence to be assessed is: "The return request for Product A has been approved and a refund may be executed." The comparison objects are the two records and the draft sentence, the period is March 2026, and the field is return request status. The two records share the same subject, period and field, and their values pending review and approved are mutually exclusive, which constitutes a contradiction. The draft sentence's claim of approval is not supported by the records, and refund execution is a business action that requires a separate downstream implementation.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the two knowledge bases that use the same vector model | The documents require the same vector model for joint association; check the model identifier one by one and exclude any that do not match |
| Search mode | Use hybrid retrieval for questions containing identifiers and status fields | Identifiers and proper nouns suit full-text retrieval, natural language questions suit semantic retrieval; compare per test question |
| Reference cap | Start at a length that covers one complete return record | The unit is tokens; test with actual materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Start at a middle value and first label relevant and irrelevant samples | The value only takes effect under semantic retrieval or result reranking; compare filtering behaviour against the labelled samples |
| Question optimization | Enable only for multi-turn follow-ups and short questions | Used for coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous and a rerank model is available | The rerank score can be used for filtering; confirm question completeness and candidate count before comparing ordering |
Three common errors
- Treating pending review as approved. The input is the two records above plus the draft sentence. Compare the values of the same field and observe that pending review and approved coexist as mutually exclusive values. The correct judgement is a contradiction, and the draft sentence is not supported.
- Treating a retrieval hit as authorization. The input is the returned reference to the return record. Compare it against the authority needed to execute a refund and observe that the reference only describes an application status and states no approval or execution authority. Record the evidence gap rather than inferring permission.
- Forcing knowledge bases with different vector models into joint association. The input is the model identifiers of the two knowledge bases. Compare the identifiers and exclude any that do not match. Do not run a cross-model search to compare missed recalls.
How to verify the setup
- Input the two record texts above and the draft sentence, and run retrieval within the configured scope. Compare whether the returned references cover both records. If the references contain the mutually exclusive status field, the condition is met; if only one record is returned, record it as incomplete recall.
- Input the draft sentence "The return request for Product A has been approved and a refund may be executed" and compare it against the status field in the record texts. Observe that the texts state no approval and no refund execution. The judgement is not supported, and no refund may be executed on this basis.
- Use a paper input of an empty reference array to check the downstream design. Compare whether the downstream continues to run with a zero-length reference. If the link is not interrupted, the condition is met. Do not claim that a fictional rule search necessarily returns an empty array.
This page addresses only TikTok Shop US store operations, cross-border materials and multilingual retrieval, and knowledge base retrieval and recall. Whether materials, operations or requests in an operations handover sit at a given authorization state cannot be settled by retrieval or by a form.