What the documents look like
Customer-service knowledge-base maintenance prepares question-and-answer material in advance and does not approve refunds directly. The routine materials are FAQ entries, reply templates, product parameter tables, and escalation tickets, drawn from the service system, product detail pages, after-sales issue statistics, and quality spot checks. For cross-border material and multilingual retrieval, the practical difficulty is that the same product carries different parameter wording across language sites, so retrieval has to match semantically close content. Semantic retrieval itself supports cross-multilingual understanding, which means a Chinese question can match an English knowledge point. Material updates run weekly or monthly; that cadence does not change the retrieval settings on this page.
A knowledge base is made of libraries, collections, and data. The smallest search unit is the library, so a search covers the whole library; collections only classify and manage data and do not affect search results. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together, so the model identifier is checked first and non-matching libraries are excluded. Retrieved reference content is output as an array and its length can be 0, which means the output path still completes even when nothing is found.
Constraints at the retrieval step
Semantic retrieval computes similarity from vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval runs both and merges the two result sets through RRF. RRF is a fusion process inside hybrid retrieval; it is not a switch that is turned on. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and it is a filter threshold rather than the final ordering rule. Retrieval and similarity scores do not carry permission or entity isolation.
A hypothetical paper sample: a customer-service knowledge-base operator is checking a multilingual FAQ entry that is being handed over. The source text reads: "FAQ-CN-021, product A, Chinese version, parameter table states 'battery capacity 4500mAh', effective date 2024-03-01, applicable site CN." A second item is a proposed handover request: "Use the reply template of FAQ-CN-021 directly for English-site retrieval, with no separate check required." The question is whether that request can be executed as written. This sample is for paper review only and does not represent a real customer or an operation that has occurred.
The material states only the Chinese version and the CN site. The English-site parameter wording is not stated, so the request is not supported by the material and the gap stays open. The request also does not state an authorizing party or an effective scope, so it cannot be treated as an approved handover.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Note and hypothetical starting point |
|---|---|---|
| Associated knowledge bases | Verify whether the selected knowledge bases use the same vector model | Libraries with different vector models cannot be associated; check the identifier and exclude non-matching ones |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | No single mode is assumed to suit all material |
| Reference limit | Unit is token | Test with material of different lengths to see whether content stays complete, then adjust the limit |
| Minimum relevance | Takes effect only in semantic retrieval or with result reranking | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was changed |
| Result reranking | Requires a ReRank model and must be enabled in the settings | Reranking returns a 0-1 score that can be used for filtering |
The reference limit is expressed in tokens rather than a top-k count, because in mixed knowledge bases the chunk lengths differ widely and a top-k result would be unstable. For this page the starting value is a hypothetical test point chosen to check whether a long FAQ entry and a short parameter line are both returned whole; it is compared against the observed completeness of the returned content, not treated as an official default.
Three common errors
- Associating knowledge bases that use different vector models. Input: two libraries whose vector-model identifiers differ. Compare the identifiers for equality. Observable result: the identifiers do not match, so the library is excluded and the actual observation is recorded, without inferring runtime behavior.
- Treating minimum relevance as a permission or entity filter. Input: one low-relevance record that belongs to another business line. Compare whether the filter changes anything other than similarity. Observable result: the similarity threshold does not distinguish business ownership, so authorization must be implemented separately.
- Treating a pending handover request as already authorized execution. Input: the proposed handover request above. Compare whether the request text states an authorizing party and an effective scope. Observable result: neither is stated, so the request is unsupported and the evidence gap is preserved.
How to verify the setup
- Input the source text "FAQ-CN-021, product A, Chinese version, parameter table states 'battery capacity 4500mAh', effective date 2024-03-01, applicable site CN" together with the proposed handover request. Compare whether the request states an authorizing party and an applicable-site scope. Observable result: neither is stated, so the item is pending verification and the handover is not executed.
- Input the same source text. Compare whether the Chinese-version parameter and the English-site retrieval need refer to the same field over the same period. Observable result: the material states only the Chinese version and the CN site; the English-site parameter is not stated, so the request is unsupported.
- Input an explicit paper empty result
[]. Compare whether the downstream design accepts a reference output of length 0. Observable result: the output path still completes, and no claim is made that a fictional rule search must return[].
These checks are paper exercises. No retrieval, deployment, or handover has been performed.