What the documents look like
Marketplace advertising adjusts bids against live traffic and conversion signals, so the working documents for this category are campaign plans, keyword reports, audience package descriptions and conversion attribution tables. Their update cadence is daily or hourly, which means a retrieval request in an order fulfilment and after-sales setting may point at material that was accurate earlier in the same day and has since been revised. The retrieval step does not resolve that tension; it returns candidate reference content, and a person or a downstream process decides what remains current.
In this task, questions typically concern the advertising source attached to one order, the progress of a refund, or a promise made in a keyword plan. Those facts sit in different documents. A campaign plan may name the keyword plan, an attribution table may link the order to that plan, and a refund record may carry its own status field. Retrieval has to bring back the passages that mention the same subject, object and state so that a reviewer can compare them.
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 search step only locates reference content. It does not judge authorization status, and it does not execute a refund or modify a campaign plan.
Constraints at the retrieval step
A single retrieval run combines several paths according to the input and the search parameters: semantic recall, full-text recall, image description or image vector recall, question optimization and reranking. The results are then merged, filtered by similarity and trimmed to a reference limit. Search modes are compared against the actual test question rather than assumed to suit all material. The reference limit is measured in tokens. The minimum relevance value applies only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Reference content is output as an array whose length may be zero, and the downstream chain still proceeds when nothing is found. The same-vector-model requirement applies to knowledge base association and does not extend to external systems. A similarity score does not perform permission or entity isolation.
Hypothetical paper exercise sample: the source material reads, "Order A-2291 submitted a refund request on 12 March 2025 and its current status is pending review; the order came from keyword plan K-08, which belongs to account Z-14." The draft sentence to be checked reads, "The refund for order A-2291 is complete, and the advertising source is keyword plan K-08." The comparison covers subject, object, state and date. The source state is pending review; the draft says complete, which is a mutually exclusive value in the same field and therefore a contradiction. The advertising source matches the source material. Pending review does not indicate that a refund has been executed, and it does not indicate that operating permission is absent.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only advertising document bases that use the same vector model | Read the vector model identifier first, exclude bases with a different identifier, and do not run a cross-model retrieval |
| Search mode | Start with hybrid retrieval | Keyword identifiers and natural-language questions appear together; compare against pure semantic and pure full-text results |
| Reference limit | Hypothetical starting point of 2000 tokens for this exercise | Test with long and short material to see whether content is complete, then compare against a smaller value for truncation |
| Minimum relevance | Hypothetical starting point of 0.3 for this exercise | Label relevant and irrelevant samples first, then compare against a higher threshold for filtering behaviour |
| Question optimization | Enabled 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 | Enabled when the candidate set is large | Rerank scores are filtered on a 0 to 1 scale; compare ordering against vector scores alone |
Three common errors
- Writing a pending-review state as complete. The input is the source sentence "its current status is pending review." The comparison takes the same order and the same field, and a draft that says "complete" is a contradiction. Executable check: extract the status field, compare it character by character against the source, and mark any difference for manual confirmation.
- Treating a retrieval hit as authorized execution. The input is a refund description returned by retrieval. The comparison asks whether that description states an approval or a completed execution. If it does not, record the gap and do not write that the refund has been paid. Executable check: inspect the reference content for an authorization or execution field, and keep the item unconfirmed when the field is absent.
- Associating knowledge bases across vector models. The input is the vector model identifier of two knowledge bases. The comparison checks whether the two identifiers match. If they differ, exclude the base and stop; do not continue into a recall comparison. Executable check: read the model identifier field and allow joint association only when the values are identical.
How to verify the setup
- Using the hypothetical sample draft sentence, input "The refund for order A-2291 is complete, and the advertising source is keyword plan K-08." Compare the source state "pending review" with the draft state "complete." A mutually exclusive value in the same field is judged a contradiction, and a matching advertising source does not offset the state contradiction.
- Using the same material, input "Order A-2291 came from keyword plan K-08 and belongs to account Z-14." Compare the subject, object and source fields. The source-check condition is met only when all three match the source material and the state has not been rewritten.
- For empty-result handling, input the paper reference array []. Compare whether the downstream design proceeds along the chain with length zero. Do not claim that a fictional rule makes retrieval return an empty array.
The public technical facts checked for this page, used to identify relevant content in the documents rather than to be copied as a fixed five-row table: the knowledge base search node can associate one or more knowledge bases that use the same vector model; search modes are compared against the test question across semantic, full-text and hybrid retrieval, and no single mode is assumed to suit all material; the reference limit is measured in tokens, and content completeness is tested with material of different lengths before the limit is adjusted; minimum relevance applies only under semantic retrieval or when result reranking is used, and thresholds are set after labelling relevant and irrelevant samples; question optimization performs coreference resolution and question expansion, and the original and optimized questions are compared to check whether the subject was altered.