What the documents look like
Procurement for marketplace advertising and on-site traffic draws on four recurring document types: campaign plans, keyword reports, audience package descriptions, and conversion attribution tables. Supplier materials arrive as quotation sheets, qualification statements, and delivery schedules. These materials update daily or hourly, and the same supplier may appear in several campaign plans with different quotation and delivery terms.
For the vector model and index step, the material that matters is the split, indexed, and recalled content that later supports citation review. These documents carry time and subject attributes. A quotation sheet states a quotation validity period; a campaign plan states a campaign period. The two do not substitute for each other. Retrieval results reflect textual similarity only. They do not determine whether a supplier is authorized, whether a quotation has taken effect, or whether a delivery schedule has been confirmed.
Two features of this category shape the task. First, advertising procurement adjusts prices against traffic and conversion in near real time, so the underlying documents change often and old and new versions can coexist. Second, the people handling this work are campaign specialists, brand marketing staff, or agency buyers, and they need to trace every recalled statement back to a specific supplier document. A retrieval hit that cannot be traced to a subject, object, and status in the source text is not usable evidence for procurement review.
Constraints at the vector indexes step
Knowledge bases joined for retrieval must use the same vector model. Before joining, check the actual model and the input type each base accepts. Bases with different models cannot be joined together. One source data block can be associated with multiple indexes, and the index content and the original content are reviewed separately. Indexes are best built from complete and focused information; retrieval coverage and index length are weighed against the sample at hand.
Image retrieval depends on whether image descriptions or image vectors were generated at ingestion time. Looking only at the query side is not enough to decide whether an image can be found. Semantic retrieval and the minimum relevance value act on similarity ordering only. They do not perform permission or entity isolation. A recalled supplier document does not mean that supplier has passed admission or received authorization.
Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a switch that can be turned on. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or result reranking. After a model or index change, compare citation source text using the same set of questions, and do not promise an outcome from a vector score alone.
The document search node can be added repeatedly, has external input, static configuration, and a trigger. It accepts one or more knowledge bases that share the same vector model. Its citation output is an array whose length can be 0, so the downstream path still runs when nothing is found.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this page's sample. The values are hypothetical starting points for this trial, used to compare against citation source text for the same question set. They are not official defaults.
| Configuration or check item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Vector model of joined knowledge bases | Record the actual model identifier of each base | Exclude bases with mismatched models; do not perform cross-model joint retrieval |
| Index content granularity | One focused index group per source data block | Compare index text against the source sentence by sentence; confirm no unrelated passage is mixed in |
| Image index | Check whether ingestion generated an image description or image vector | Record separately from query-side upload behavior; if either is absent, mark as unknown |
| Citation limit | Set to a token value the sample can hold | Compare against total returned citation length; adjust when exceeded |
| Minimum relevance | Set to one trial value between 0 and 1 | Compare filtering effect only under semantic retrieval or result reranking |
| Result review method | Compare citation source text using the same question set | Observe whether citations contain the supplier and quotation fields present in the sample |
Three common errors
- Treating a retrieval hit as authorization confirmation. Recalling a supplier's quotation sheet shows textual similarity only. Whether the quotation has taken effect and whether the supplier has been admitted still require separate review. Check: input the sample quotation source text, compare whether the returned citation contains the quotation validity field; if the citation shows only an amount without a validity period, record it as an evidence gap.
- Joining knowledge bases across different vector models. When bases use different vector models, joining them and comparing recall on that basis is unsupported. Check: input each base's model identifier, compare whether the identifiers match, exclude the mismatched base, and record the actual observation without inferring missed recall or redundancy.
- Replacing field review with a semantic score. Using the minimum relevance score to decide whether supplier materials are complete is unsupported. Check: input the sample draft sentence, compare whether the citation source text contains subject, object, and status fields; when the score is high but fields are missing, record it as unsupported.
How to verify the setup
The verification below uses the hypothetical trial sample from earlier. The sample source text is: "Hypothetical trial sample: Supplier A submitted an inventory procurement quotation sheet on the hypothetical date 1 March 2025, with a quotation validity period ending 31 March 2025, and a delivery schedule pending confirmation." The draft sentence to be judged is: "Supplier A has completed inventory procurement quotation confirmation, and the delivery schedule has taken effect." The subject, object, and status in this draft sentence must each be compared against the source text.
- Input the draft sentence and run retrieval. Compare whether the returned citation contains the quotation validity period and the delivery schedule status. If the citation shows only a quotation amount without the validity period and schedule fields, judge it as unsupported and record the evidence gap.
- Input the vector model identifier of each joined base. Compare whether the identifiers match. If a mismatch exists, exclude that base and record the actual observation. Do not perform cross-model retrieval to compare missed recall.
- Input a short question containing only the supplier name. Compare whether the returned citation is an empty array. If an empty array is returned, review how the downstream design carries a citation path of length 0. If content is returned, check whether the citation actually contains that supplier field. Do not guarantee a first hit or full recall.