What the documents look like
Short-video affiliate sales distributes products through content that stays visible over time and through product links placed in that content, so the work does not depend on live explanation during a stream. The material handled day to day includes short-video scripts, lists of products carried by those videos, commission plans, and violation feedback. In store service and internal training, staff answer questions about which product a video carries, how long a commission plan runs, and what a violation notice requires, using the versions of these materials that have already been stored in a knowledge base. Update frequency is daily and weekly, and a retrieval result reflects content as of the moment it was stored. A retrieval hit therefore does not make an authorization effective during a store handover.
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. Before execution, check the vector model identifier of each base, exclude those that do not match, and record what was observed. Do not run a cross-model retrieval to compare missed recalls. Search mode is compared across semantic retrieval, full-text retrieval, and hybrid retrieval for the test question, and no single mode is assumed to suit all material. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization handles coreference resolution and question expansion. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process and not a standalone switch. Reranking and image retrieval require the corresponding model capability and must be enabled in settings. Optional paths shown in the documented flow diagram do not mean every request executes all of them. Retrieval and relevance scores do not carry out permission or entity isolation, and a retrieval hit does not mean an operation has been authorized.
Hypothetical drill sample: a store's internal training material reads, "The commission plan for carried product A takes effect on March 1, 2025, and remains valid through June 30, 2025; store staff may introduce this product according to the script during the validity period." The handover request to be assessed is, "Please confirm that the commission plan for product A can still be used for store introduction in July 2025." The subject of this request is a store staff member, the object is the commission plan for product A, the state is a use request after the validity period has ended, and the effective range stops at June 30, 2025.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases holding scripts, product lists, commission plans, and violation feedback whose vector model identifiers match | Check the vector model identifier of each base; exclude any whose identifier does not match |
| Search mode | Start with hybrid retrieval | Compare recalled content against semantic retrieval and full-text retrieval on the same test question, and observe differences between identifier-type and natural-language questions |
| Citation limit | Hypothetical starting point of 800 tokens | Test with actual material of different lengths to see whether cited content is complete, and compare truncation against longer or shorter values |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then compare filtering results against higher or lower thresholds |
| Question optimization | Enable first in multi-turn follow-up and short-question scenarios | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enable according to settings and observe scores | Compare candidate ordering before and after reranking and check whether scores are used for filtering |
Three common errors
- Treating a retrieval hit as authorized execution. In the sample, the commission plan for product A stops being valid on June 30, 2025. Retrieving that material shows only that the content was stored; it does not show that the plan can still be used in July. The input to check is the material text and the handover request, the comparison is between the validity field and the request date, and an observed date beyond the validity period is recorded as pending verification rather than written as a denial or a conflict.
- Associating knowledge bases that use different vector models. Check the vector model identifier of each base, exclude those that do not match, and record the actual observation. Do not infer missed recalls or redundancy from this.
- Treating optional retrieval paths as executing on every request. Image retrieval requires an image input and the corresponding model capability, and reranking requires the model and must be enabled in settings. When the matching input or capability is absent, record that the path did not execute and do not infer its returned result.
How to verify the setup
- Input the material text for the commission plan of product A from the hypothetical sample together with the handover request, compare the validity field against the request date, and observe whether the result falls inside the validity period. When it falls outside, record it as pending verification and keep the evidence gap open.
- Input the same test question under semantic retrieval and under full-text retrieval, compare the recalled content from each against identifier-type and natural-language phrasing, and observe whether one path returns nothing. When nothing is returned, record the actual observation without guaranteeing a first hit or full recall.
- Input an explicit paper-level empty result
[]to check the downstream design, and observe whether the output chain still runs when the cited content length is 0. This check does not claim that a search against a fictional rule will necessarily return[], and it does not claim that an unrelated query must be empty.