What the documents look like
Concession-counter documents centre on joint-operation percentage contracts, mall event letters, sales receipts and monthly reconciliation statements. Their sources are the mall POS, the brand ERP, mall membership campaigns and the finance system. Because the mall collects payment centrally and settles on a percentage, the sales evidence and the settlement basis sit in two separate systems. One transaction can therefore appear with different fields on the mall receipt and in the brand ERP.
Cross-border material and multilingual retrieval adds a further layer: contracts and event letters may exist in more than one language version. The same entity, the same period and the same field must be aligned across languages before any conclusion can be drawn. A percentage rate stated in one language version and a settlement period stated in another are not automatically the same fact.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together. The node outputs cited content as an array, and that array can have a length of zero, so the output path still runs when nothing is retrieved.
Semantic search relies on vector distance and suits natural-language questions and cross-language understanding. Full-text search relies on keyword hits and suits numbers, model designations and proper nouns. Hybrid search merges the routes through RRF, and the documentation describes RRF as a fusion step rather than a switch to be turned on. Reranking requires a rerank model and must be enabled through the settings. Image retrieval requires an image input and the corresponding model capability; it is an optional path in the overall flow and does not run on every request.
The minimum relevance value is a filter threshold that takes effect only in semantic search or when result reranking is used. It does not perform permission or entity isolation. Retrieval is not authorisation: signing, approval and settlement all require a separate downstream implementation. The same-model requirement for linked knowledge bases applies to the knowledge bases in this node and does not extend to external systems.
Hypothetical paper exercise. The source material to be checked reads: "The concession joint-operation percentage contract (hypothetical reference H-01) provides that the party's monthly sales data is based on the mall POS summary, and the percentage rate follows the contract annex. The mall event letter (hypothetical reference H-02) states that the event period is the hypothetical 1 March to 31 March." The draft sentence to be assessed reads: "The concession's March percentage fee has been settled in full at the percentage rate confirmed in the event letter."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Basis and comparison method |
|---|---|---|
| Linked knowledge bases | Link only two bases that share one vector model | Check the vector model identifier and exclude bases whose identifier differs |
| Search mode | Try semantic search and full-text search separately first | Compare the content recalled by each route for the same test question |
| Citation limit | Hypothetical starting point of 2000 tokens | Check content completeness against actual material of different lengths, then adjust |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then compare the filtered results |
| Question optimisation | Left off for now | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Left off for now | Confirm the rerank model is available, then enable it through the settings |
The citation limit is expressed in tokens, not in a count of items. The starting values above are chosen for this exercise so that there is a concrete value to compare against; they are not documented defaults. The comparison method matters more than the number: a limit is adjusted after observing whether long and short source documents both survive intact.
Three common errors
- Treating a retrieval hit as business confirmation. Input: the contract text under hypothetical reference H-01 and the draft sentence. Compare whether the cited content contains evidence of completed settlement. The observed result is that the citation covers the percentage rate and the data source only, and the claim of completed settlement is not supported.
- Treating a pending state as authorisation. Input: the event letter text under hypothetical reference H-02. Compare the event period stated in the letter with the settlement period named in the draft sentence. The observed result is that the event period covers March, while the letter does not state a confirmed percentage rate, so that field remains unverified.
- Linking knowledge bases that use different vector models. Input: the vector model identifiers of the two bases. Compare whether the identifiers match. The observed result is that they differ, so the non-matching base is excluded and the observation is recorded. No cross-model retrieval is executed.
How to verify the setup
- Input the contract text under hypothetical reference H-01 together with the draft sentence "The concession's March percentage fee has been settled in full at the percentage rate confirmed in the event letter." Compare whether the cited content contains evidence of completed settlement and of a confirmed percentage rate. The condition is met only when content matching the draft sentence's claim of completed settlement appears; otherwise record the claim as unsupported.
- Input the event letter text under hypothetical reference H-02. Compare whether the event period stated in the letter and the period named in the draft sentence are the same field for the same entity. A contradiction is judged only when the same entity, the same period and the same field carry clearly mutually exclusive values. When the field is not stated, record it as unverified.
- Input the vector model identifiers of the two knowledge bases. Compare whether the identifiers match. Bases may be linked together only when the identifiers match; when they differ, exclude the non-matching base and record the observation. Do not infer missed recalls or redundant returns from this check.
- Input an explicit paper empty result of
[]and check whether the downstream design still runs the output path when the citation length is zero. Do not claim that any particular retrieval rule is guaranteed to return[].
A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Where the source material does not state a field, the gap remains a gap.