What the documents look like
Seller ratings and experience score maintenance revolves around consumer feedback that arrives after fulfilment and after-sales service. The working materials are negative review handling records, experience score detail statements, service diagnostic reports, and rectification ledgers. Sources include Douyin Store experience scores, Tmall service indicators, JD after-sales reviews, and customer service systems, with daily or weekly updates. In store service and internal training, these materials answer frontline questions about handling standards, rectification requirements, and supplementary evidence.
Three facts shape retrieval for this task. First, the materials mix long natural-language descriptions with precise identifiers. Negative review reasons are narrative, while order numbers, indicator names, and rectification ticket numbers are exact strings; a single retrieval mode rarely covers both. Second, the same rectification ledger repeats across store, period, and indicator dimensions, and chunk lengths vary widely, so a token-based citation limit behaves differently on long and short chunks. Third, training questions often arrive as follow-up turns with missing subjects, which raises the question of whether question optimization should be enabled.
Constraints at the retrieval step
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. The handling step is to check each base's vector model identifier, exclude those that do not match, and record what was observed; running a cross-model search to compare missed recalls is not part of the procedure.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval relies on keyword hits and suits identifiers and proper nouns. Hybrid retrieval uses both paths and merges them through RRF. RRF is a fusion process, not a switch that can be turned on. Result reranking uses a ReRank model to reorder candidates and produces a score from 0 to 1 that can be used for filtering. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule, and it does not perform permission or entity isolation. Retrieval scope is determined by the associated knowledge bases, and retrieval itself does not grant access authorization to records.
The citation limit is measured in tokens and controls how much content a single search may cite. The node outputs citations as an array, and the length may be 0; when nothing is found, the output path still completes. Question optimization runs before retrieval, performs coreference resolution and question expansion, and adds one model call. Image retrieval requires an image in the input and a model with the corresponding capability, and reranking requires a model and must be enabled by configuration. These optional paths do not mean every request passes through all of them.
Configuration and implementation checks
The hypothetical drill sample below illustrates configuration and checking method. It does not represent any real store, real record, or completed test.
Hypothetical drill sample: in a store's rectification ledger for the period March 1 to March 31, one record reads "Rectification ticket ZG-0312, Store A, indicator is after-sales response time, rectification statement submitted March 12, status pending review." In the same period's negative review handling records, one entry reads "March 12 order reported delivery delay, consumer contacted for supplementary evidence, evidence not returned." The handover request to be judged is "Synchronize the rectification conclusion of ZG-0312 to the training team for use as an effective standard."
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the rectification ledger and negative review handling records | First check that both bases share the same vector model identifier and exclude mismatches |
| Search mode | Hybrid retrieval | Compare against semantic-only and full-text-only results on identifier-type questions |
| Citation limit | Set to 2000 tokens for this hypothetical trial | Test with long and short actual materials for truncation, then compare against a smaller value |
| Minimum relevance | Set to 0.3 for this hypothetical trial | Label relevant and irrelevant samples first, then compare filtering against a higher threshold |
| Question optimization | Enable only for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a model is available | Compare candidate ordering before and after, and check whether the score can be used for filtering |
Three common errors
- Treating a retrieval result as authorization. In the sample, ZG-0312 has status pending review. Retrieving that record shows the content was recalled; it does not show the rectification was approved or the standard took effect. Pending review does not indicate missing operational permission, and it does not indicate approval.
- Using minimum relevance as scope isolation. Minimum relevance is a 0-to-1 filter threshold that takes effect only under semantic retrieval or result reranking. It cannot be used to conclude that unrelated stores or unauthorized records were excluded.
- Treating optional paths as always executed. Image retrieval requires an image input and model capability, and reranking requires a model and must be enabled by configuration. When conditions are not met, that path does not run, and the flow diagram does not support inferring that every request passes through every step.
How to verify the setup
- Check the vector model identifiers of the associated bases. Input: the vector model identifiers of the two bases in the hypothetical sample. Compare whether the two identifiers match. A match satisfies the condition for joint association; when they do not match, exclude the non-conforming base and record the observation without inferring runtime results of cross-model association.
- Check whether the cited content is complete. Input: the ZG-0312 record text from the hypothetical sample and the current citation limit. Compare whether the recalled content contains all five items: rectification ticket number, store, indicator, date, and status. All five present and not truncated satisfies the condition; missing items are recorded as evidence gaps.
- Check whether the proposed sentence is supported. Input: the handover request to be judged and the hypothetical sample text. Compare the same subject, the same period, and the same field. The source status field reads pending review, and the request states the standard is effective; the two values are mutually exclusive on the same field, so the request is not supported. If the source does not state an effective status, record it as pending verification.
Empty-result branch: use an explicit paper input of an empty array to check the downstream design and confirm the output path still completes. Do not claim that a fabricated-rule search necessarily returns an empty array.