What the documents look like
Marketplace advertising adjusts bids against traffic and conversion in near real time, so the citation work behind product and marketing copy runs on four recurring documents: the delivery plan, the keyword report, the audience package description, and the conversion attribution table. The public brief describes data refresh at a daily and hourly level, with source verification completed before copy is finalized and delivery execution beginning afterward. A promotional sentence that drifts away from its source material keeps consuming budget on the delivery side, and the attribution table cannot reconcile the actual transaction basis. For citation tracing, this means every claim in a draft sentence must map back to a specific passage in one of those four documents, with the same subject, the same period, and the same field.
The knowledge base search node connects one or more knowledge bases that share the same vector model. Its output is an array of citations whose length may be zero; an empty array still completes the output path and does not mean the workflow terminated with an error. Retrieval output passes to the answering step, where each answer sentence is checked against the cited source text. The minimum relevance value acts as a filter threshold only during semantic retrieval or result reranking, and it does not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, so final ordering does not strictly follow a single vector similarity score. Image retrieval requires image input and model capability, and reranking requires a model that is enabled according to settings. Document timeliness is checked by comparing the current version against a historical version, and whether a citation is adopted still depends on valid-version review.
Constraints at the citation tracing step
Hypothetical paper exercise sample. The source material reads: "The final price shown on this product page applies to orders placed between March 1, 2025 and March 31, 2025; after that date the new page governs." It also reads: "Gift quantity is governed by what the order checkout page displays; gift images on the main page are illustrative only." The draft sentence awaiting judgment is: "Order in March and receive a double gift; the final price continues through the end of the month." The verification subject is the same product, the same campaign period, the same price field, and the same gift field. The draft sentence states the gift quantity as double, while the source material only says the checkout page governs. The draft sentence states the final price as continuing through the end of the month, while the source material specifies that orders placed by March 31 qualify. Both points are unsupported by the source; neither is a contradiction.
A contradiction requires the same subject, the same period, and the same field with explicitly mutually exclusive values. Where the source is silent, or where only evidence from a related stage exists, the result is unsupported or pending verification. A pending review does not demonstrate a lack of operating permission. Signing, reaching cooperation, and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. The materials needed for verification must sit inside a target retrieval scope the reviewer can access, or inside a clearly stated manual input; when they are missing, the gap is recorded as a gap.
Configuration and implementation checks
| Configuration item | Assumed starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One delivery copy knowledge base and one product material knowledge base | First check that both vector model identifiers match; exclude any that differ and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Compare recalled content for the same draft sentence against semantic retrieval and full-text retrieval separately |
| Citation limit | Set to 6000 tokens as an assumed starting value for this exercise | Compare against the length of source paragraphs needed for the draft sentence, and adjust to cover complete paragraphs |
| Minimum relevance | Set to 0.5 as an assumed starting value for this exercise | Compare against rerank scores and observe whether candidates below that value are filtered |
| Question optimization | Enabled | Compare recalled content for the same short question against the disabled state |
| Citation output handling | Retrieval output feeds the answering step | Check that each answer sentence maps to cited source text, and check downstream design when the array is empty |
The citation limit uses tokens as its unit, corresponding to the instability of top k chunk lengths. The minimum relevance figure is an assumed starting value for this exercise, not an official default.
Three common errors
- Treating minimum relevance as a permission or entity isolation mechanism. The value is a filter threshold that applies only during semantic retrieval or result reranking. It does not exclude unrelated businesses or unauthorized records. A retrieval action does not constitute authorization.
- Describing RRF as an independent switch that can be turned on. RRF is the fusion process for multiple result paths inside hybrid retrieval and relates to the search mode selection; no separate switch item exists.
- Moving thinking budget into the citation token limit. Thinking budget controls model reasoning intensity and is a different field from the citation limit; the two cannot be interchanged.
How to verify the setup
- Input the hypothetical sample's source material and draft sentence, then compare the gift quantity and price period in the draft sentence against the corresponding source fields. Both points in the draft sentence are unsupported and are recorded as pending verification, not as contradictions.
- Input the same draft sentence and run semantic retrieval, full-text retrieval, and hybrid retrieval separately, then compare the recalled content. Observe whether hybrid retrieval merges multiple result paths through RRF. First-position hits and complete recall are not guaranteed.
- Input a paper-form empty array as the citation output and check whether the downstream answering step is designed to handle an empty result. An empty array means no content was retrieved; the output path completes and does not indicate that the workflow terminated with an error.
The document flow diagram lists optional paths that do not all execute on every request: image retrieval requires corresponding image input and model capability, and reranking requires an available model enabled according to settings. Knowledge bases built on different vector models cannot be associated together. Check the identifiers first, exclude those that do not match, and record actual observations rather than describing runtime outcomes for a cross-model association.