What the documents look like
Procurement and supplier material for marketplace advertising and on-site traffic sits alongside campaign plans, keyword reports, audience package descriptions, and conversion attribution tables. Those operational files update daily or hourly. The procurement side adds supplier quotations, purchase quantities, delivery batches, and settlement terms. The values in those fields determine whether a campaign budget can be committed.
Retrieval works on text fragments drawn from this material. Vector distance measures semantic closeness. Exact identifiers such as quotation numbers, batch numbers, and purchase order numbers are better served by full-text retrieval. The choice between the two paths follows the content of the material, and one mode cannot be assumed to cover every document.
Constraints at the retrieval step
A knowledge base is organized into library, collection, and data. The smallest search unit is the library. Collections classify data for management and do not affect search results. 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 node outputs referenced content as an array, and the length can be zero. When nothing is retrieved, the output path still completes. Semantic retrieval, full-text retrieval, and hybrid retrieval use different recall paths. Hybrid retrieval merges multiple result sets through RRF, then applies similarity filtering and a citation ceiling to produce the referenced content. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it acts as a filter threshold rather than a final ordering rule. Question optimization performs coreference resolution and question expansion, and it adds an extra model call. Image retrieval requires image input and corresponding model capability. Reranking requires a model and must be enabled in configuration. Retrieval itself performs no authorization judgment. Approval status, signature status, and contract effectiveness in procurement must be checked separately.
Hypothetical paper exercise: the source material reads, "Supplier A submitted a quotation on 10 March 2025, quotation number CG-20250310-07, valid until 10 April 2025, purchase quantity 12,000 units, delivery in 3 batches." The handover request under review reads, "Please confirm this quotation is effective and arrange the first 4,000 units for warehouse receipt." The source states a quotation validity period. It states no supply term, no contract effectiveness, and no warehouse receipt authorization. The phrases "is effective" and "arrange for warehouse receipt" are not supported and remain pending verification.
Configuration and implementation checks
The settings below relate to this page's sample. They list documented search parameters or implementation checks. The values are starting points for this hypothetical trial, used for comparison against other values, and are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Include the library holding this supplier's quotation material | Check whether the vector model identifier matches across libraries, and exclude those that do not |
| Search mode | Hybrid retrieval | The quotation number needs full-text hits and delivery descriptions need semantic recall; compare single-path and hybrid recall |
| Citation ceiling | Set to hold the quotation number and delivery batch wording in full | Compare material fragments of different lengths to see whether referenced content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then set a value | Takes effect only under semantic retrieval or result reranking; compare candidates before and after filtering |
| Question optimization | Enable for short questions, disable for questions containing explicit numbers | Compare the original question with the optimized question, and check whether the subject or number was altered |
| Result reranking | Enable when candidates are numerous | The rerank score supports filtering; compare ordering before and after reranking |
Three common errors
- Treating a retrieval hit as granted authorization. In the sample, quotation number CG-20250310-07 being recalled only shows the material falls inside the retrieval scope. A validity date of 10 April 2025 does not state a supply term or warehouse receipt authorization. For the check, enter that number, compare the recalled content against the phrase "is effective," and observe whether the material states any effectiveness clause.
- Associating libraries built on different vector models. For the check, enter the vector model identifier for each library, compare whether they match, exclude those that do not, and record the actual observation. Do not infer runtime behavior of a combined association.
- Treating minimum relevance as an entity or permission filter. For the check, enter quotation material from the same supplier and material from an unrelated supplier, compare candidates before and after filtering, and observe whether unrelated material can still be recalled on semantic closeness. This value performs no permission isolation.
How to verify the setup
- Enter the sample source text and the request "Please confirm this quotation is effective and arrange the first 4,000 units for warehouse receipt." Compare whether the recalled content contains the quotation validity period and delivery batches, and observe whether "is effective" has a corresponding statement in the material. When the material does not state it, record the item as pending verification.
- Enter quotation number CG-20250310-07 and run full-text retrieval and semantic retrieval separately. Compare the recalled content from both paths and observe whether the number is matched exactly. If the full-text path returns no hit, record that observation without inferring a cause.
- Enter a paper request that clearly has no corresponding material, compare the output reference array, and observe whether the length is zero and whether the downstream path completes. Handling of an empty result follows the downstream design. Do not claim that a fictional rule necessarily returns an empty array.