What the documents look like
Rating maintenance materials center on consumer feedback that follows after-sales fulfillment. The main sources are Douyin store experience scores, Tmall service indicators, JD after-sales reviews, and customer service systems. Everyday documents include negative review handling records, experience score details, service diagnostic reports, and rectification ledgers. Updates arrive daily and weekly.
When these materials enter a knowledge base, the retrieval step has to handle the same negative review appearing under inconsistent field definitions across the customer service system, the experience score details, and the rectification ledger. Attribution of a negative review, collection of supporting evidence, and tracking of rectification all depend on comparing fields across documents. Whether a retrieval result covers multiple fields for the same object determines whether a review conclusion holds up.
A knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. One group of data can map to multiple vectors, and the vector score takes the highest score. A single retrieval combines question optimization, semantic recall, full-text recall, image handling, RRF fusion, and ReRank reordering according to the input and search parameters, then applies similarity filtering and a citation limit before outputting an array of cited content, which may have a length of zero.
Constraints at the retrieval step
Semantic retrieval relies on vector distance. Full-text retrieval relies on keyword matches. Hybrid retrieval merges both through the RRF formula. The minimum relevance value takes effect only during semantic retrieval or when result reordering is used, and it acts as a filter threshold; it does not perform permission or entity isolation. Linked knowledge bases must use the same vector model, and libraries with different models cannot be linked together. When different identifiers appear, verify them first and exclude those that do not match. Do not run cross-model retrieval to compare missed recalls.
Image retrieval requires image input and the corresponding model capability. Reordering requires a model and must be enabled by configuration. Both are optional paths and do not execute on every request.
Hypothetical paper exercise sample: the source material reads, "In the experience score details for Store A from March 1 to March 7, 2025, the after-sales review item is recorded as 4.6 points; the negative review handling record for the same period states that a negative review on March 3 is in a pending supporting evidence status; the rectification ledger states that the responsible stage for this negative review is logistics and the rectification status is tracking in progress." The draft sentence to be judged is, "Store A's after-sales review item scored 4.6 points during this period, and the negative review has been fully rectified." This draft sentence conflicts with the tracking in progress status in the rectification ledger and is unsupported. The handover request reads, "Operations is asked to confirm the evidence supplement result for this negative review before March 10." This request is in a pending confirmation state and does not constitute an authorization conclusion.
Configuration and implementation checks
The table below lists only officially documented settings and implementation checks relevant to this page's primary task. The values are starting points for this hypothetical exercise, not official default values.
| Configuration item | Starting point and reason for this hypothetical exercise | Comparison method and observation target |
|---|---|---|
| Linked knowledge bases | Select the negative review handling record library and the rectification ledger library that use the same vector model | Check whether the vector model identifiers of each library match, and exclude those that do not |
| Search mode | Test natural language questions with semantic retrieval first | Compare hit content item by item against full-text retrieval and hybrid retrieval results |
| Citation limit | Start at 2000 tokens because the sample contains long records with multiple fields | Test with materials of different lengths to see whether cited content is complete |
| Minimum relevance | Start at 0.4 and first label relevant and irrelevant samples | Compare how labeled samples are retained before and after the threshold |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reordering | Enable when there are many candidates | Compare ordering before and after reordering and observe whether scores are more precise |
Three common errors
- Treating minimum relevance as a permission or entity isolation measure. The input is a cross-store retrieval request, the comparison object is the threshold setting and the returned records, and the observed result is that adjusting the threshold does not change the visibility of unauthorized records. When authorization is missing, record the gap.
- Linking knowledge bases that use different vector models. The input is two libraries with different model identifiers, the comparison object is the identifier field, and the observed result is that they cannot be linked together. Exclude those that do not match and record the actual observation without inferring runtime results.
- Reading a pending supporting evidence status as completed rectification. The input is the negative review handling record and the rectification ledger source text, the comparison object is the status field of the same negative review, and the observed result is that the two status values differ, so the draft sentence conclusion is unsupported.
How to verify the setup
- For the draft sentence "Store A's after-sales review item scored 4.6 points during this period, and the negative review has been fully rectified," the input is the experience score details and the rectification ledger source text, the comparison object is the rectification status field of the same negative review in the same period, and the observed result is that the ledger records tracking in progress, so the rectification completion conclusion in the draft sentence does not hold and must remain pending verification.
- For the handover request "Operations is asked to confirm the evidence supplement result for this negative review before March 10," the input is this request text and the negative review handling record, the comparison object is the evidence status field, and the observed result is that the record shows pending supporting evidence and the request remains pending confirmation, which does not constitute approval or execution.
- For the empty result branch, the input is a paper-constructed citation content array [], the comparison object is the downstream node's design for accepting this empty array, and the observed result is that the output link still proceeds. Do not claim that a fictional rule retrieval will necessarily return an empty array.