What the documents look like
Marketplace advertising adjusts bids against live traffic and conversion, so the material handled in operating reviews and anomaly post-mortems differs from promotion enrollment files. The working documents are delivery plans, keyword reports, audience package descriptions and conversion attribution tables. A delivery plan carries budget and bidding entities. A keyword report carries term and spend fields. An audience package description carries targeting scope. A conversion attribution table carries the attribution basis. Retrieval terms frequently contain identifiers, model numbers and proper nouns, which makes full-text retrieval and hybrid retrieval relevant for this material. Retrieval returns quoted content; it does not grant any authorization.
Update frequency for this category can be daily or hourly, but that cadence does not change the retrieval step described here. What matters at this step is that the source documents hold identifiers and status fields that must be read exactly as written.
Constraints at the retrieval step
A 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 vector model identifier is checked and non-matching bases are excluded. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a switch. The reference limit is measured in tokens. The minimum relevance threshold takes effect only under semantic retrieval or when result re-ranking is used. Question optimization performs coreference resolution and question expansion, and it adds an additional model call before retrieval. Quoted content is output as an array whose length can be zero, and the downstream path still runs when nothing is retrieved. Image retrieval and re-ranking are optional paths that require the corresponding input and model capability.
Hypothetical paper exercise sample: on 10 June 2025, a brand marketing department delivery specialist checks a paper document titled "Qianchuan delivery plan 0620." The document text reads: "Plan number QC-0620; delivery entity is a brand flagship store; daily budget and bid range are in the attached schedule; execution status is pending review; the plan may go live only after review approval." A separate draft handover request reads: "This plan has been approved; please execute delivery against the daily budget and synchronize the keyword report." The retrieval scope is limited to this delivery knowledge base, and the vector model identifier is A. The points to be determined are whether the request's statement "has been approved" agrees with the document's "pending review," and whether executing delivery is an action this retrieval step can carry.
Configuration and implementation checks
| Configuration item | Setting or check method | Relation to this page's sample |
|---|---|---|
| Associated knowledge bases | Check the vector model identifier; associate only knowledge bases with the same vector model | The sample fixes vector model A; bases with a different identifier are excluded |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | The sample contains identifier QC-0620, so full-text and hybrid results can be compared |
| Reference limit | Unit is tokens; test completeness with material of differing lengths | This hypothetical trial start value is chosen to cover the full plan text, and is compared against a shorter value to see whether content stays complete |
| Minimum relevance | Takes effect only under semantic retrieval or result re-ranking | The sample begins with semantic retrieval; relevant and irrelevant samples are labelled before adjusting the threshold |
| Question optimization | Compare the original question with the optimized question and check the subject | The sample question contains a subject and an identifier; check whether optimization alters the subject |
| Result re-ranking | Used when a re-ranking model is available and enabled by configuration | When the sample returns many candidates, check the re-ranking score and filtering |
Three common errors
- Treating a retrieved "pending review" document as evidence of approval. Input is the sample document text and the draft request; the comparison is "pending review" against "has been approved"; the observable result is that the two values are mutually exclusive on the same field, so the request is not supported.
- Treating retrieval or a form as authorization to execute. Input is the draft request phrase "execute delivery and synchronize"; the comparison is the retrieval step's output against a business execution action; the observable result is that retrieval returns quoted content only, and execution requires a separate downstream implementation.
- Associating knowledge bases with different vector models. Input is the vector model identifiers of two bases; the comparison is whether the identifiers match; the observable result is that the non-matching base is excluded and no cross-model retrieval is performed.
How to verify the setup
- Take the sample document text and the draft request, enter the retrieval scope, and compare the document's "pending review" against the request's "has been approved." If retrieval returns that document text and the two values are mutually exclusive, the request is not supported. If the document text is not returned, record the evidence gap.
- Take the sample plan number QC-0620 and run full-text retrieval and hybrid retrieval separately. Compare whether the returned quoted content contains the content corresponding to that identifier. If it does, record a hit. If it does not, record the observation without concluding that all content was recalled or that the result must be empty.
- Take an explicit paper empty input and check the downstream design for handling a quoted-content array of length zero. If the downstream handles the empty-array branch, the condition is met. Do not claim that a fabricated rule retrieval necessarily returns an empty array.
The same-field comparison rule applies throughout: a determination requires the same subject, the same period and the same field, with mutually exclusive values on that field. A status of pending review does not by itself show a lack of operating permission, and an unsigned termination is not a completed termination. Where a field is not stated, or only a related stage is evidenced, the result remains unsupported or pending verification rather than being recorded as a contradiction.