What the documents look like
Short-video affiliate sales relies on long-term content exposure and product-link conversion. A published video keeps generating plays and transactions after release, without a live host explaining the item. The working documents for this category are short-video scripts, product lists attached to the cart, commission plans, and violation feedback. These materials update at daily or weekly frequency.
The comparison task places four kinds of material in one context: product wording in a script, product entries in a cart list, commission-rate terms in a plan, and handling records in violation feedback. The operator checks whether the subject, the object, the status, and the valid range agree across these materials. Because several materials must be held in context at once, token consumption decides which parts can be quoted in full.
Constraints at the context limits step
The knowledge-base reference limit is measured in tokens. It does not use top k, because chunk lengths differ widely in mixed knowledge bases and top k results become unstable. A model's maximum context is the length of content a single request can reference. The reply limit controls only the length of the answer. These two settings serve different purposes and cannot substitute for each other.
Hybrid retrieval combines multiple result paths through RRF. RRF is a fusion step inside the retrieval flow, not a switch that can be turned on. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used. It does not isolate business entities or unauthorized records, and retrieval itself does not grant authorization.
Hypothetical paper exercise. Source text one: "Party A commissions Party B to publish cart promotion content on a short-video platform. The product is a certain model of insulated cup. The commission rate is settled at fifteen percent of the transaction amount. The cooperation period runs from March 1, 2025 to May 31, 2025." Source text two: "On April 10, 2025, Party B submitted a cart product list. The listed products are the insulated cup and a cup brush. The cup brush falls outside the commissioned scope." Draft sentence to be judged: "Party B has completed commission settlement for all cart products, and Party A needs no further verification." The draft sentence receives no support from the source texts on product scope or on settlement status.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| AI model maximum context | Check against actual model capability | Compare with total material length to confirm the content needed for comparison is not truncated |
| Reply limit | Keep enough length to state the basis and the handling action | Compare with the number of fields the draft sentence requires |
| Knowledge-base reference limit | Set a trial starting point in tokens | Compare with single-section material length and the question sample, and observe whether content is clipped |
| Search mode | Use semantic retrieval and full-text retrieval together | Check precise fields such as product model and dates with full-text retrieval |
| Minimum relevance | Check only under semantic retrieval or result reranking | This value does not exclude unrelated businesses or unauthorized records |
| History input | Compare the same question with empty history and with history | Rule out comparison deviation caused by context differences |
A fixed reply format, when used, requires verifying that fields are complete and parsing succeeds. Checking text length alone does not confirm the format works. The knowledge base is searched as a whole; collections only classify data and do not affect search results. Database internals such as the vector field or the indexes array are implementation details of the storage layer and are not configuration items for this comparison task.
Three common errors
- Treating the reply limit as the full context window. The reply limit controls answer length; the maximum context controls the length a single request can reference. Mixing them lets material be truncated without notice. The check is to input source text one and source text two, then compare the content visible to the model against the total length of the source texts.
- Treating minimum relevance as a permission filter. This value only filters low-relevance results. It does not separate business entities or unauthorized records. The check is to input the list containing the insulated cup and the cup brush, observe whether both kinds of entry enter the candidate set, and confirm that filtering happens only by relevance.
- Writing a pending state as completed settlement. The draft sentence states that commission settlement for all cart products is complete. The source texts record only the cooperation period and the commission rate, and do not record settlement completion. The check is to compare the draft sentence with the source texts field by field and confirm that the settlement status field receives no support.
A missing record stays missing. A pending review does not show that an operation permission is absent. Signing, reaching cooperation, and contract effectiveness are three separate facts. The comparison must stay within the same subject, the same period, and the same field. Only mutually exclusive values in the same field count as a contradiction. A field that is not recorded, or that has only related-stage evidence, is unsupported or pending verification.
How to verify the setup
- Input source text one and source text two. Compare the content visible to the model against the total length of the source texts, and confirm that the product scope, the cooperation period, and the commission rate all appear in full with no truncation. The expectation is that all three sections can be checked. If one section is missing, adjust the reference limit and repeat.
- Input the draft sentence and the source texts. Compare product scope and settlement status field by field, and confirm that the cup brush falls outside the commissioned scope and that settlement status is not recorded. The expectation is that the draft sentence fails on both product scope and settlement status. If the model outputs "settlement completed," record that as unsupported.
- Input the same comparison question with empty history and with history. Compare whether the two outputs cite the same materials, and confirm that history input does not introduce conclusions beyond the source texts. The expectation is that both cite the same range. If the run with history adds a settlement conclusion not recorded in the source texts, record the context difference.
For an empty-result case, a paper input of [] can be used to check the downstream design. A relevance score is not a category, entity, or permission filter. Retrieval does not grant authorization. Cross-model association is not possible; check identifiers first and exclude those that do not match, rather than running a cross-model search to compare missed recalls. The optional paths in the overall retrieval flow are not all executed on every request: image retrieval requires image input and model capability, and reranking requires a model and must be enabled by setting. When a required material is missing from the accessible search scope or from a defined manual input, record the gap rather than filling it with an assumed authorization or effectiveness clause.