What the documents look like
This category centres on disposable plastic product reporting and environmental supervision. The working materials are plastic bag procurement ledgers, sales vouchers, semi-annual reports and self-discipline commitments. Sources include the commerce ministry reporting system, POS plastic bag sales, the procurement system and store ledgers. In store service and internal training, frontline staff answer questions about ledger field meanings, voucher retention requirements and reporting scope using these materials. Update frequency covers both semi-annual and daily records, so retrieval configuration has to accommodate periodic submission documents alongside day-by-day sales records at the same time.
The consequences of error include deadline-ordered rectification, fines, and regulatory interviews. When a retrieval result is treated as an authorisation basis, the risk moves from information deviation into compliance handling. Retrieval and relevance scoring do not perform permission or entity isolation. Finding content does not mean authorisation has been granted.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. At retrieval, vectors are recalled first, then the vector ID is used to look up the original data in MongoDB dataset.datas. When the same original data is hit by multiple routes, the results are merged and the highest score is taken. Semantic retrieval, full-text retrieval and hybrid retrieval are selected according to question type. Hybrid retrieval is merged through RRF, and reranking produces a 0-1 score used for filtering. The reference cap is measured in tokens. The minimum relevance value takes effect only during semantic retrieval or when result reranking is used. Retrieval output is an array whose length can be 0, and the empty-result path still completes.
A hypothetical paper exercise sample: a store ledger entry reads "On 12 March 2025, this store purchased 400 vest-style plastic bags, supplier Company A, purchase order CG-2025-0312, received into inventory." A draft training sentence reads "This store completed all plastic bag procurement and inventory receipt for March 2025 and can provide a compliance certificate externally." A pending handover request reads "Please submit the semi-annual report to the regulator based on the above ledger."
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the plastic bag ledger base and the report base that share one vector model | Check the vector model identifier first, exclude bases with a different model, do not run cross-model retrieval |
| Search mode | Run one round of semantic retrieval and one round of full-text retrieval | Compare the reference content of the two modes using the same question |
| Reference cap | Set to a token volume covering one complete ledger record | Compare against the ledger source length and confirm the record is not truncated |
| Minimum relevance | Mark relevant and irrelevant samples first, then set a threshold | Compare against the marked results and observe whether irrelevant fragments are filtered |
| Question optimisation | Enable for follow-up questions containing references, disable for explicit questions | Compare the original question and the optimised question, check whether the subject was changed incorrectly |
| Result reranking | Enable when candidates are numerous and observe the score | Compare against the unranked order and confirm the relevance score can be used for filtering |
Three common errors
- Treating a retrieval hit as authorised execution. In the hypothetical sample, the draft sentence states "can provide a compliance certificate externally", while the ledger source only records procurement and inventory receipt and does not state any external provision authority. Compare the ledger fields against the scope claimed by the draft sentence, and observe whether authorisation wording appears that the source does not contain.
- Associating knowledge bases across vector models. When bases with different vector models are associated together, compare each base identifier and observe whether a base with an inconsistent model has been included. The documents do not describe the runtime result of cross-model association; only the identifier check and exclusion action are recorded.
- Mixing different periods or fields to judge a contradiction. The hypothetical ledger is a single entry dated 12 March 2025, while the draft sentence states "completed all procurement and inventory receipt for March 2025". Compare the date range and the record count, and observe whether a single entry can support a whole-month conclusion. Matters not stated are recorded as pending verification.
How to verify the setup
- Input the hypothetical ledger source and the draft sentence, compare the "completed all procurement and inventory receipt" claim against the single ledger entry, and observe whether the draft sentence exceeds the source scope. If the draft sentence keeps the source date, quantity and order number and does not extend the conclusion, the condition is met.
- Input the hypothetical handover request "Please submit the semi-annual report to the regulator based on the above ledger", compare the requested action against what the ledger source records, and observe whether the ledger contains submission authorisation and semi-annual report fields. If the request is marked as pending authorisation verification, the condition is met.
- Input an empty reference array
[]to the downstream, compare whether the downstream design handles the empty-result path, and observe whether any assumption depends on reference content being non-empty. If the downstream continues with the empty array and does not claim retrieval always returns content, the condition is met.