What the documents look like
Short-video affiliate sales depend on long-term content exposure and product-link conversion, without live explanation. The main data sources are the selected affiliate program, short-video playback data, product transaction data and platform rules. The documents handled day to day are short-video scripts, linked-product lists, commission plans and violation feedback. In operating reports and exception reviews, these materials serve as the basis for cross-checking, and the retrieval step has to return the subject, object, status and validity scope from scripts, lists, plans and feedback accurately enough for a person to compare.
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 using different vector models cannot be associated together. The output is an array of references, and its length can be 0, which means the output path still runs even when no content is found. Semantic retrieval, full-text retrieval and hybrid retrieval are compared against test questions; no single mode is presumed suitable for all materials. The minimum relevance value takes effect only in semantic retrieval or when result re-ranking is used. Question optimization performs coreference resolution and question expansion, and the subject has to be checked for whether it was altered incorrectly. Retrieval and similarity scores do not carry authorization or entity-isolation duties.
Hypothetical drill sample: the source text reads "From 1 March 2025 to 31 March 2025, influencer A's linked video V-017 uses commission plan P-203, and the commission rate follows the value stated in the plan; plan P-203 was in pending-review status on 20 March." The draft sentence to be judged reads "Influencer A completed commission settlement under P-203 for the whole of March." This request involves checking the same subject, the same period and the same field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the script base and the plan base that use the same vector model | Verify that the vector model identifiers match before associating; exclude those that do not match |
| Search mode | Test the draft sentence containing the identifier under full-text retrieval first | Identifier and model-number questions suit full-text retrieval; then compare with semantic retrieval results |
| Reference limit | Start at a token amount that covers one complete plan clause | Test with actual materials of different lengths to see whether content is complete, then adjust |
| Minimum relevance | Start at a middle value, effective only in semantic retrieval or re-ranking | Label relevant and irrelevant samples first, then adjust the threshold and compare results before and after filtering |
| Question optimization | Enable for short questions containing coreference | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result re-ranking | Enable when a re-ranking model is available and enabled by configuration | Compare candidate order before and after re-ranking and observe whether the score is used for filtering |
Three common errors
- Treating the retrieval score as a permission filter. The input is influencer A's draft sentence, the comparison object is the returned reference content, and if plan clauses of other influencers are returned, retrieval did not perform entity isolation and a person has to exclude them.
- Treating pending-review status as already effective. The input is the source text stating "in pending-review status on 20 March," the comparison object is the draft sentence "completed commission settlement under P-203 for the whole of March," and if the draft does not mark the pending-review scope, it is judged unsupported.
- Forcing knowledge bases with different vector models into one association. The input is two knowledge bases with different vector model identifiers, the comparison object is the node association configuration, and if the identifiers do not match while association is still attempted, the mismatched base should be excluded and the actual observation recorded.
How to verify the setup
- Input the source text and the draft sentence, and compare whether the draft sentence has the same subject, the same period and the same field as the source. If the draft writes a pending-review status as completed settlement, judge it unsupported and keep the evidence gap.
- Input the draft sentence containing the identifier and observe the returned references under full-text retrieval and semantic retrieval separately. If full-text retrieval matches the plan identifier while semantic retrieval does not, record the difference between the two modes; a first-position match is not guaranteed.
- Input a clear paper-based empty result [] to check the downstream design, and observe whether the output path runs when the reference-content array length is 0. If downstream still requires continued processing, record that branch and check the downstream design; do not claim that a fictional rule retrieval necessarily returns an empty array.