What the documents look like
Ordering events serve offline dealer reservations and do not directly manage platform storefronts. Inventory procurement and supplier material typically sits in four document types: ordering handbooks, ordering contracts, sample lists, and credit approvals. Brand channel departments, regional dealers, and merchandise planning teams use these materials to confirm order quantities, supply arrangements, and payment terms. Materials refresh on a quarterly or annual cycle, so retrieval configuration can be reviewed against an annual or quarterly version.
The four document types carry different retrieval needs. Sample lists contain model numbers, codes, and proper nouns, where full-text retrieval is more likely to match. Ordering handbooks contain policy descriptions and natural-language explanations, where semantic retrieval fits better. Credit approvals and ordering contracts involve parties, amounts, and validity scopes, so complete quotation of the source text matters; a similarity score alone cannot decide these fields.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. At operation time, confirm the identifier first and exclude those that do not match; do not run cross-model retrieval to compare missed recalls. Retrieval runs against the whole base. Collections exist only for classification management and do not affect search results.
The reference cap is measured in tokens, not top k. Minimum relevance is a 0-1 filter threshold that takes effect only under semantic retrieval or when result reranking is used. It does not perform permission or entity isolation, and it cannot be used to exclude unrelated business lines or unauthorized records. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process, not a switch that can be turned on. Reranking requires a model and must be enabled according to settings. Image retrieval requires the corresponding image input and model capability. The optional paths shown in the documented flow diagram do not mean every request executes all of them. The output reference content is returned as an array and its length can be 0.
Hypothetical paper exercise sample: a brand channel department plans to check a handover request for one sample list from the 2025 spring ordering event. The source material reads: "Sample code SP-2025-031, item name autumn-winter windbreaker, suggested retail price 899 yuan, minimum order quantity 50 pieces, supply cycle 45 days after order confirmation." The draft sentence to be judged: "The supply cycle for SP-2025-031 is 30 days after order confirmation." The source and the draft differ in the supply cycle field for the same subject and the same field, an explicit mutual exclusion, so the draft is judged contradictory.
Configuration and implementation checks
| Configuration item | Value or check action on this page | Basis and notes |
|---|---|---|
| Associated knowledge base | Select the inventory procurement and supplier material base | Must use the same vector model as other bases; confirm the identifier first |
| Search mode | Use full-text or hybrid retrieval for model and code questions | Semantic retrieval is unstable for proper nouns |
| Reference cap | Hypothetical trial starting value 3000 tokens | Chosen because sample list entries are long; compare content completeness against 1500 tokens |
| Minimum relevance | Hypothetical trial starting value 0.4 | Label relevant and irrelevant samples first; compare false recalls against 0.6 |
| Question optimization | Enable for multi-turn follow-ups with unclear references | Performs coreference resolution and question expansion; check whether the subject is altered |
| Result reranking | Enable when candidates are numerous and the question is complete | Requires a ReRank model; enable according to settings |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input one query, compare whether the returned references contain other dealers or unauthorized records, and observe whether out-of-scope content still appears. The relevance score carries no isolation duty; out-of-scope access depends on downstream permission implementation.
- Forcing association of bases on different vector models. Input two bases with different vector model identifiers, compare whether the identifiers match, and observe whether the mismatched base is excluded. The documentation states only that the same vector model is required; it does not describe runtime results after forced association.
- Treating a pending credit approval as approved. Input one credit approval record, compare whether its status field reads approved, and observe whether the draft sentence writes it as orderable. Pending review does not demonstrate a lack of operating permission, and it does not equal approval.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence "The supply cycle for SP-2025-031 is 30 days after order confirmation," compare the supply cycle field values, and observe whether the retrieval result recalls the source and shows 45 days. Recalling the source with a matching field indicates that this entry can be retrieved.
- Input a query containing only the model number "SP-2025-031," compare the returns from full-text and hybrid retrieval, and observe whether the sample entry is matched. A match indicates that code-type queries can use the full-text path; a miss is recorded as pending verification and is not judged as the data not existing.
- Input an explicit paper input of an empty array [], compare whether downstream nodes continue with the empty reference, and observe whether the output chain breaks. An output length of 0 with the chain continuing indicates that empty-result handling matches the documented behavior; do not claim that a fictional rule retrieval always returns [].