What the documents look like
Channel receivables collection happens after credit sales. The work centres on credit granting and cash recovery rather than on matching buyers with sellers. Records come from finance systems, bank receipts, distributor orders and sales contracts. The documents handled day to day are the receivables ageing table, the dunning letter, the credit agreement and the shipment freeze notice. Order fulfilment and after-sales question answering require checking subject, object, status and validity scope across these materials. The ageing table shows how overdue balances are distributed. The dunning letter shows which demands have already been issued. The credit agreement shows the limit and its term. The shipment freeze notice shows what has been done to the shipment step. Materials update weekly or monthly. The retrieval function itself does not depend on that rhythm, but whether quoted content still represents the current state needs human judgement.
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 built on different vector models cannot be associated together. At operation time, check the vector model identifier first, exclude those that do not match, and record what was actually observed. Semantic retrieval relies on vector distance. Full-text retrieval relies on keyword hits. Hybrid retrieval fuses the two paths through RRF. RRF is a fusion process, not a switch that can be turned on. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. It is a filter threshold and does not perform permission or entity isolation. The citation limit is measured in tokens and controls how much content one search may quote. Question optimisation handles coreference resolution and question expansion, and it adds an extra model call. Image retrieval requires image input and the corresponding model capability. Re-ranking requires a model and is enabled according to settings. Citation content is output as an array, and its length may be 0.
Hypothetical paper exercise sample: a brand finance staff member holds three materials. Material A is an excerpt from a credit agreement: "Party A grants Party B a credit limit of RMB 500,000, with a credit term from 1 January 2025 to 31 December 2025." Material B is an excerpt from a shipment freeze notice: "From 15 October 2025, shipments to Party B beyond the credit limit are suspended." Material C is an excerpt from a dunning letter: "As of 30 September 2025, Party B owes RMB 180,000 for goods." The draft sentence to be assessed: "Party B's credit term has expired, so all shipments should stop." This sentence concerns the subject Party B, the object credit term, and the status expired. Material A states the credit term runs to 31 December 2025, while the draft sentence says it has expired. The two take mutually exclusive values on the same field. Material B states that shipments beyond the limit are suspended, while the draft sentence says all shipments stop. The scopes differ. The materials needed for the assessment must sit within a target retrieval scope the reviewer has access to, or within an explicit manual input. When they are missing, record the gap.
Configuration and implementation checks
Configuration centres on the checking needs of this sample. The numbers are starting points for this hypothetical exercise, used to compare citation completeness across different values. They are not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one base that uses the same vector model | Check the vector model identifier first, exclude those that differ, record the actual observation |
| Search mode | Hybrid retrieval | Compare citation differences for the same question against semantic retrieval and full-text retrieval separately |
| Citation limit | Hypothetical starting point of 2000 tokens | Compare whether content stays complete using materials of different lengths, then adjust |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, then compare citation changes after filtering |
| Question optimisation | Enabled | Compare the original question with the optimised question, and check whether the subject was altered |
| Result re-ranking | Enabled according to settings | Confirm the re-ranking model is available, and compare ordering before and after re-ranking |
Three common errors
- Treating a retrieval result as the basis for authorisation. The input is the draft sentence "Party B's credit term has expired, so all shipments should stop." The comparison object is the term field in Material A. Acting to stop shipments on that basis turns retrieved content into an authorisation to execute. Executable check: input the draft sentence, compare the term stated in Material A with the term wording in the draft sentence, observe whether the two take mutually exclusive values on the same field, and record the result as pending verification.
- Associating knowledge bases that use different vector models. The input is two knowledge base identifiers. The comparison object is each base's vector model identifier. Attempting association when the identifiers differ is cross-model association. Executable check: check the vector model identifiers of the two knowledge bases, observe whether they match, exclude and record those that do not, and do not run a cross-model retrieval comparison.
- Treating a pending-review status as approved. The input is a credit agreement marked "pending review." The comparison object is the status field in the agreement. Concluding that the limit is already effective on that basis is a status misreading. Executable check: read the status wording in the original agreement, observe whether any approval or effectiveness record exists, and record the result as unsupported when it is missing.
How to verify the setup
- Input the original text of Material A and the draft sentence "Party B's credit term has expired, so all shipments should stop." Compare the term field in Material A with the term wording in the draft sentence. Observe whether the two take mutually exclusive values on the same field. If they do, the pending-verification condition is met, and the evidence gap is recorded.
- Input the original text of Material B and the draft sentence "all shipments should stop." Compare the freeze scope in Material B with the stop scope in the draft sentence. Observe whether the two scopes match. If they do not, the pending-verification condition is met, and the scope difference is recorded.
- Input an empty array as the citation content, and compare how the downstream design handles an empty result. Observe whether the downstream branches on the empty result. If the downstream design explicitly handles an empty result, the condition is met. Retrieval checks record observations and failure branches only. They do not guarantee a first-hit match, full recall, or that an unrelated query must return empty.