What the documents look like
This category covers disposable plastic bags and single-use items reporting for supermarkets and convenience stores, where the work is organised around environmental reporting rather than membership promotion. The documents handled day to day are plastic bag procurement ledgers, sales vouchers, half-year reports and self-discipline commitments. These materials come from different systems: the Ministry of Commerce reporting system, POS plastic bag sales, the procurement system and store ledgers. They also belong to different entities and different periods. A procurement ledger records supplier, incoming quantity and date. A sales voucher records retail activity at one store. A half-year report aggregates usage and recovery information. Retrieval therefore faces several sources with different field definitions, and each item has to be checked by entity, object, period and status. Sales quantity cannot be treated as procurement quantity, and one store's ledger cannot be treated as a head-office aggregate.
Constraints at the retrieval step
A knowledge base search node can be associated with one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the vector model identifier must be checked first and non-matching bases excluded, with the actual observation recorded. The search mode can be semantic retrieval, full-text retrieval or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process, not a separate switch that can be turned on. The citation limit controls how much content each search cites, measured in tokens. Minimum relevance takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule, and it does not perform permission or entity isolation. Question optimisation performs coreference resolution and question expansion and adds an extra model call. Cited content is returned as an array whose length can be zero, and the downstream path still runs when nothing is found. Image retrieval requires corresponding image input and model capability, and reranking requires a rerank model enabled according to configuration. Optional paths in the overall flow do not mean every request executes all of them. Retrieval itself does not constitute authorisation; forms collect values, and business verification, approval, database writes and external synchronisation each need their own downstream implementation.
Hypothetical paper exercise sample: the source material reads, "Party A, a hypothetical supermarket chain company, and Party B, a hypothetical plastic products factory, signed an annual supply letter of intent on a hypothetical date, agreeing that Party B supplies according to Party A's monthly orders, and that an order takes effect after written confirmation by Party A's procurement manager; this letter of intent does not constitute a minimum purchase commitment." The draft sentence to be assessed is: "Party B has obtained full supply authorisation for the current year, and Party A must accept goods monthly." The check points are whether the letter of intent has been converted into an effective order, whether there is written confirmation by the procurement manager, and whether the authorisation scope covers all supply. The sample contains no confirmation record and no order number, so the draft sentence is unsupported and remains pending verification.
Configuration and implementation checks
| Configuration item | Value or check on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Associate only the procurement ledger base and supplier material base whose vector model identifiers match | The node requires the same vector model; different models are not associated together |
| Search mode | Hypothetical trial starting point for this exercise is hybrid retrieval, compared with full-text retrieval results | Observe full-text hit differences for queries such as numbers and supplier names |
| Citation limit | Hypothetical trial starting point for this exercise is 2000 tokens, compared with 1000-token results | Unit is tokens; observe citation completeness against ledger paragraph length |
| Minimum relevance | Hypothetical trial starting point for this exercise is 0.3, compared with 0.5 | Takes effect only in semantic retrieval or reranking; label relevant and irrelevant samples first |
| Question optimisation | Enabled for multi-turn follow-up scenarios, disabled as a control for single clear questions | Used for coreference resolution and question expansion; adds one model call |
| Result reranking | Enabled according to configuration; check whether a rerank model is available | Rerank score is 0-1 and can be used for filtering |
Three common errors
- Treating a letter of intent as an effective supply authorisation. The sample source text states only that an order takes effect after written confirmation by the procurement manager, and no confirmation record is present, while the draft sentence claims full supply authorisation has been obtained, which is unsupported. Check action: input the draft sentence and the source text, compare the authorisation status field, and observe whether a written confirmation record appears; when no record exists, mark it as pending verification.
- Treating minimum relevance as a permission filter. Minimum relevance only filters low-relevance results and does not exclude unauthorised records. Check action: input a query involving another supplier, compare the returned content with the access permission list, and observe whether that record is still returned; if it is returned, the filter does not perform isolation.
- Treating RRF as a switch that can be turned on. RRF is the result fusion process of hybrid retrieval, and the documentation provides no separate switch. Check action: check whether the configuration interface contains such a switch field, compare the actual configurable items with the documented description, and record the missing field as a non-configurable item when it is absent.
How to verify the setup
- Authorisation status check. Input the hypothetical sample source text and the draft sentence, compare the authorisation status and written confirmation fields, and observe whether the authorisation claimed by the draft sentence has a corresponding record in the source text; when no record exists, output pending verification rather than a determination of authorisation.
- Empty-result handling check. Input a paper empty array [] as the cited content, compare whether downstream nodes run normally, and observe whether the path breaks because of the empty result; the documentation states the length can be zero and the path still runs.
- Cross-model association check. Input two knowledge bases whose vector model identifiers differ, compare whether the identifiers match, and observe whether they are excluded from joint association; record only the identifier check result, and do not run a cross-model retrieval to compare missed recalls.
- Question optimisation subject check. Input a multi-turn question containing a reference, compare the optimised question with the source subject, and observe whether the subject is altered; when the subject changes, record the deviation and do not infer a retrieval failure cause from it.
- Citation limit completeness check. Input ledger paragraphs of different lengths, compare the cited content under different limits, and observe whether long paragraphs are truncated; when truncation occurs, record the actual observation and do not state that a value is an official default.