What the documents look like
Customer-service knowledge maintenance works with FAQ entries, reply templates, product parameter tables and escalation tickets. Customer-service supervisors, knowledge-base operators and product operations maintain these materials together, and the update rhythm is weekly or monthly. When the material is wrong, the visible results are answers that miss the question, escalated complaints and failed customer-service training.
This page covers one slice of operations authority and handover management: the retrieval and recall step. It describes which materials, actions and requests sit in which authorization state during a handover, and it treats retrieval and form entry as content collection or content return. Authorization decisions require a separate check. The knowledge-base 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. During handover, the vector model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. A cross-model retrieval is not run in order to compare missed recalls. The search node returns cited content as an array, and the array length can be 0.
Constraints at the retrieval step
Retrieval does not perform authorization decisions. Semantic recall and relevance scores are content-similarity measures. They do not act as a category, entity or permission filter, and they do not exclude unrelated business lines or unauthorized records. A relevance score is not a category, entity or permission filter. The minimum relevance value takes effect only in semantic retrieval or when result re-ranking is used, and it is a filter threshold rather than a final ordering rule. The citation limit is measured in tokens and controls how much content a single search may cite. Question optimization runs before retrieval, performs coreference resolution and question expansion, and adds one model call. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process inside hybrid retrieval and is not described as a standalone switch. Re-ranking requires a model and must be enabled in the settings. Image retrieval requires an image input and the corresponding model capability. These optional paths do not all run on every request. An empty citation output can pass through the downstream chain. During handover, a clearly labelled paper input of [] is used to check the downstream design. No claim is made that a fictional rule makes retrieval always return [].
Configuration and implementation checks
The table below is built around a hypothetical drill sample and lists documented settings or implementation checks that relate to this page's feature and primary task.
| Configuration or check | Starting point for this hypothetical drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Two FAQ bases, both labelled with the same vector model identifier | Check the identifier before associating; compare with a cross-model base and exclude non-matching ones |
| Search mode | Hybrid retrieval | Compare with semantic retrieval and full-text retrieval results, and observe the hit difference on numbered questions |
| Citation limit | Hypothetical starting value of 800 tokens | Test content completeness with FAQ entries of different lengths, and compare with a smaller value to see whether truncation appears |
| Minimum relevance | Hypothetical starting value of 0.4 | Label relevant and irrelevant samples first, then compare the filtering effect with a higher or lower value |
| Question optimization | Enabled | Compare the original question with the optimized question, and check whether the subject was changed incorrectly |
| Handover check action | Manually record the vector model identifier and the authorization state | Record this separately from the retrieval return, and do not use retrieval results as authorization evidence |
Three common errors
- Treating a retrieval hit as an authorization pass. The input is one retrieval request and one pending handover request for FAQ rewrite permission. Compare the cited content returned by retrieval with the authorization field. If permission is judged as granted on the basis of the cited content alone, record this as an error, and the authorization state remains pending verification.
- Associating cross-model knowledge bases directly. The input is two knowledge bases labelled with different vector model identifiers. Compare whether the identifiers match. If the association is executed and recall differences are compared on that basis, record this as an error. The correct action is to exclude the non-matching base and record the actual observation.
- Treating minimum relevance as a permission or entity filter. The input is one unrelated business query and one query for an unauthorized record. Compare the returned results with the permission field. If the relevance score is used to judge that a record has been excluded, record this as an error. The score does not carry that function.
How to verify the setup
Hypothetical drill sample: the source material text is "FAQ-101 entry: a return request must be submitted within seven days of signed receipt; customer service may register it on the customer's behalf, and approval is completed by the after-sales supervisor." and "FAQ-102 entry: product parameters follow the current display on the detail page; the knowledge-base parameter table is for internal checking only." The draft sentence to be judged is "Customer service has obtained the return approval permission for FAQ-101 and can approve refunds directly." Judgment conditions: same subject, same period, same field. The source text states only that customer service may register on the customer's behalf, and that approval is completed by the after-sales supervisor. The approval permission described in the draft sentence is not supported and remains pending verification. It does not constitute a conflict with the source text.
- Input the source material text above and the draft sentence. Compare whether "register on the customer's behalf" and "approve refunds directly" are mutually exclusive values in the same field. If the permission described in the draft sentence has no corresponding statement in the source text, judge it as unsupported and keep the evidence gap open.
- Input two knowledge bases labelled with different vector model identifiers and one association request. Compare whether the identifiers match. If they do not match, exclude the association, record the actual observation, and do not run a cross-model retrieval to compare missed recalls.
- Input a paper empty result of [] and the downstream node design. Compare whether the downstream chain passes through with the empty citation. If the downstream chain accepts it normally, record that this design is usable. No claim is made that a fictional rule makes retrieval always return [].