What the documents look like
Customer-service knowledge base maintenance covers FAQ entries, script templates, product parameter tables, and escalation tickets. The work is front-loaded reference material preparation; it does not approve refunds directly. Sources include the customer-service system, product detail pages, after-sales issue statistics, and quality inspection sampling. Update frequency is weekly or monthly, which matters here because a comparison run depends on whether the relevant FAQ text has already been stored and indexed at the time of the request.
The comparison task takes a draft sentence or handover request and checks it field by field against stored source text: subject, object, status, and validity range. Retrieval decides whether comparable source text can be found at all. When no source text is retrieved, the judgment stays at an evidence gap and cannot supply authorization or effective clauses that the source does not contain.
Constraints at the retrieval step
A knowledge base is organized as library, collection, and data; the smallest search unit is the library. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries using different vector models cannot be associated together, so identifiers should be checked first and non-matching libraries excluded.
Semantic retrieval relies on vector distance. Full-text retrieval suits numbers, model codes, and proper nouns. Hybrid retrieval merges multiple result paths through RRF. Result reranking uses a ReRank model for a second ordering. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a final sorting rule, and it does not perform permission or entity isolation. The reference limit is measured in tokens. Reference content is output as an array and its length can be zero.
Hypothetical paper sample: a knowledge base contains one FAQ source text reading "The filter replacement cycle for the A-brand X1 water purifier is 6 months, counted from the day after signed receipt. This description applies to complete-machine orders sold on or after January 1, 2025." A separate handover request draft reads: "Please arrange a filter replacement reminder for A-brand X1 water purifiers sold in December 2024 according to the above cycle." The check compares whether the sale date falls inside the validity range stated in the source text. This sample is a paper exercise and does not refer to any real customer or order.
Configuration and implementation checks
| Configuration item | Hypothetical starting value and reason for this page | What to compare against |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model library containing the FAQ source text above | Check whether the vector model identifiers of the libraries match |
| Search mode | Test the sample question separately under semantic retrieval and full-text retrieval | Compare whether the source text is retrieved under each mode |
| Reference limit | Hypothetical starting value covering the tokens needed for the complete source sentence | Compare with the actual length of the source text to confirm the content is not truncated |
| Minimum relevance | Hypothetical starting value at a middle point to observe the filtering effect | Compare against samples labelled relevant and irrelevant |
| Question optimization | Enable for questions containing references; disable for the sample question with a complete subject | Compare whether the question subject is altered incorrectly before and after optimization |
| Result reranking | Enable according to whether a ReRank model is available | Compare the rerank score against the minimum relevance threshold |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the draft requests a reminder for orders sold in December 2024, while the source validity range is limited to orders sold on or after January 1, 2025. Retrieving the source text only shows that comparable material was found; it does not show the request has been authorized for execution.
- Treating minimum relevance as a permission or entity filter. This parameter takes effect only in semantic retrieval or result reranking, where it filters low-relevance results. It cannot be used to exclude unrelated businesses or unauthorized records.
- Treating an unstated field as a contradiction. The sample source states only the replacement cycle and validity range; it does not state the reminder method or the executing person. When the draft does not involve those fields, the record should note unsupported or pending verification, not negation or conflict.
How to verify the setup
- Input the hypothetical sample source text and draft sentence, and run retrieval with only the same-vector-model library associated. Compare whether the retrieved content includes that FAQ source text. If the reference content is an empty array, record it as not retrieved, and do not infer a system error or an automatic business action.
- Input the draft sentence and compare the sale date field against the validity range field in the source text. A contradiction is judged only when the same subject, same period, and same field carry clearly mutually exclusive values. In this sample, December 2024 and on-or-after January 1, 2025 are mutually exclusive, so the judgment is that the request falls outside the source validity range, while the reminder method and executing person fields remain missing.
- Input a follow-up question containing a reference, once with question optimization enabled and once disabled, and compare whether the optimized question subject still points to the A-brand X1 water purifier. If the subject is changed to another model, record it as the question optimization altering the subject incorrectly, and do not use that retrieval result for judgment.
- For image retrieval and result reranking paths, observe only when the corresponding image input, model capability, and enabled setting exist. The optional paths listed in the overall document flow diagram do not mean every request executes all of them. An empty-result handover can be checked against an explicit paper input of [] for downstream design, and no claim can be made that retrieval necessarily returns an empty array.