What the documents look like
Short-video affiliate sales depend on long-term content exposure and product-link conversion, not on live explanation. Product selection questions therefore draw on four document types: short-video scripts, product lists for linked carts, commission plans, and violation feedback. Scripts describe selling points and spoken structure. Product lists record product identifiers and link status. Commission plans record rates and validity periods. Violation feedback records platform handling and rectification requirements.
These materials update daily or weekly. The same product may be described differently across documents, so retrieval must support comparison rather than assume one consistent record. The supplied brief identifies matching videos to products, setting commission rates, and checking title violations as the most time-consuming tasks. Retrieval accuracy therefore affects whether promotion stays effective, whether commission is wasted, and whether a video is removed.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; collections only classify data and do not affect search results. A single retrieval combines problem optimization, semantic retrieval, full-text retrieval or hybrid retrieval, image handling, RRF fusion, and Rerank according to the input and search parameters. The final output is an array of references whose length can be zero, meaning the downstream path still executes when nothing is found.
Semantic retrieval suits natural-language questions. Full-text retrieval suits identifiers, model numbers, and proper nouns. Hybrid retrieval merges both paths through RRF. Rerank produces a score between 0 and 1 that can be used for filtering. The minimum relevance value applies only during semantic retrieval or when result reranking is used. Retrieval does not perform permission or entity isolation; a retrieved item does not grant operational authorization.
Hypothetical paper exercise. A short-video affiliate project team retains three source sentences. "Linked-cart product list (March 2026 version): Product A, link status listed, applicable commission plan Plan A." "Commission Plan A: commission rate 15%, valid until 31 March 2026." "Violation feedback record: the short-video title for Product A was flagged on 10 March 2026 as containing a prohibited term, rectification pending." A handover request awaiting review states: "Please confirm that Product A currently has a 15% commission rate and that the title violation has been fully resolved, so promotion can continue." The source text records only the link status in the list, the rate and validity period of Plan A, and that the title violation is pending rectification. It does not state that rectification is complete, and it does not state that the request has been approved.
Configuration and implementation checks
| Configuration or check item | Value or action in this exercise | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Select bases containing scripts, product lists, commission plans, and violation feedback with identical vector model identifiers | The node requires the same vector model; bases with different identifiers cannot be associated together. Check identifiers and exclude mismatches. |
| Search mode | Compare semantic, full-text, and hybrid retrieval separately using test questions | No single mode is assumed suitable for all materials; observe full-text hits for identifier-type questions. |
| Reference limit | Set a hypothetical trial starting value of 800 tokens, then compare against actual materials of different lengths | Unit is tokens; the starting value is a trial value only, used to observe whether content is truncated. |
| Minimum relevance | Label relevant and irrelevant samples first, then compare returned sets as the threshold changes | Applies only during semantic retrieval or when result reranking is used; the threshold is a filter condition and does not perform permission filtering. |
| Problem optimization | Compare original and optimized questions for short queries such as "it" or "this plan" | Used for coreference resolution and question expansion; verify whether the subject was altered incorrectly. |
| Result reranking | Observe second-pass ranking scores only when a rerank model is available and enabled by configuration | Reranking is an optional path; when not enabled, this path does not execute. |
Three common errors
- Treating a retrieval hit as authorization. Input "Product A commission rate 15%" and retrieve the Plan A source text. The comparison object is the validity period in Plan A against the current date. The observed result is that the source text states validity until 31 March 2026. If the current date falls outside that range, the rate value is still retrieved, but this must not be used to conclude that promotion can continue. Retrieval does not perform authorization decisions.
- Reading pending rectification as completed handling. Input the violation feedback source text "rectification pending." The comparison object is the rectification status field for the same product and same period. The observed result is that the source text does not state rectification is complete. When completion evidence is missing, record it as pending verification. Do not conclude that it is handled or unhandled.
- Forcing association across vector models. Input two knowledge bases with different vector model identifiers. The comparison object is the vector model identifier field of each base. The observed result is that the identifiers differ. Record only that the joint association condition is not met. Do not execute cross-model retrieval, and do not infer missed recalls or redundant results.
How to verify the setup
- For the source text "Commission Plan A: commission rate 15%, valid until 31 March 2026," input the question "What is Product A's commission rate and validity period?" Compare whether the retrieved content contains both the rate and validity period fields. The condition is met when both source sentences appear in the reference content. If only the rate appears, record the missing field.
- For the pending handover request "Please confirm that Product A currently has a 15% commission rate and that the title violation has been fully resolved," input this request and compare the returned references against the status field in the violation feedback source text. The condition is met when "rectification pending" appears in the references and no completion statement appears. Do not output a conclusion that rectification is fully resolved.
- For empty-result continuation, input the paper input
[]and compare whether the downstream node continues to execute with an array length of zero. The condition is met when the downstream path still runs. Do not claim that a specific rule retrieval necessarily returns an empty array. - For cross-model bases, input two knowledge bases with different vector model identifiers and compare the identifier fields. The condition is met when the identifiers differ and the pair is recorded as not jointly associable. Do not execute a retrieval comparison.
- For minimum relevance, input labeled relevant and irrelevant samples and compare the returned sets before and after threshold adjustment. The condition is met when relevant samples are retained and irrelevant samples are reduced. The threshold is not used to exclude unauthorized records.