What the documents look like
Channel receivables collection happens after credit sales. The work centres on credit usage and cash recovery rather than on matching buyers with sellers. The main data sources are the finance system, bank receipts, distributor orders and sales contracts. Day-to-day materials include the receivables ageing table, dunning letters, credit agreements and shipment freeze notices. Ageing tables and payment records are typically updated by week or by month.
These materials carry amounts, periods, counterparties and authorisation states at the same time. A retrieval step that returns a passage does not decide whether a shipment freeze was approved, whether a credit agreement is still valid, or whether a collector is allowed to act on a given distributor account. The retrieval output has to hand the full original text to the downstream judgement, and the judgement itself stays a separate step.
For operations handover, this matters because the person taking over needs to know which materials exist, which states are recorded in them, and which states are still open. The retrieval layer can surface the right documents, but it cannot supply an authorisation conclusion, an approval, or a business action that has not been recorded.
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. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. Re-ranking requires a model and has to be enabled in the settings. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used.
Reference content is output as an array, and its length can be zero. When no content is found, that output path still completes. A similarity score supports ordering and filtering. It does not isolate counterparties, permissions or business states. A relevance threshold is not a category, entity or permission filter, and retrieval is not authorisation.
Hypothetical paper test case. Source text one: "Distributor A in the East China region, credit agreement number HX-2024-017, credit limit RMB 800,000, payment term 60 days, agreement signing date 1 March 2024, valid until 28 February 2025." Source text two: "Distributor A's ageing table for November 2024 shows RMB 120,000 overdue by 31 to 60 days and RMB 50,000 overdue by 61 to 90 days." Source text three: "On 3 December 2024 a sales support staff member submitted a shipment freeze application; the application status is pending review and the approver's comment field is empty." Draft sentence to be assessed: "Distributor A's credit agreement has expired and the shipment freeze has been executed." Handover request: "The receiving person is asked to confirm the freeze status from the above materials and continue collection."
The draft sentence contains two claims. The first claim concerns the agreement validity period. The second claim concerns the freeze status. These are separate facts recorded in separate fields. A pending application is not an approval. An offered option is not execution. An unsigned termination is not a completed termination.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this page | Reason and comparison method |
|---|---|---|
| Linked knowledge base | One collection-materials knowledge base using the same vector model | Check the vector model identifier first; exclude knowledge bases whose identifier differs, and do not run cross-model joint retrieval |
| Search mode | Hybrid retrieval as the hypothetical test starting point | Ageing tables contain numbers and proper nouns while dunning letters contain natural language; compare semantic, full-text and hybrid recall paths |
| Reference cap | 1,500 tokens as the hypothetical test starting point | Test whether long and short materials are truncated; compare completeness against the values 1,000 and 2,000 |
| Minimum relevance | 0.4 as the hypothetical test starting point | Label relevant and irrelevant samples first, then compare filtering behaviour against 0.3 and 0.5 |
| Question optimisation | Enabled as the hypothetical test starting point | The handover request contains a reference; compare the original question and the optimised question to check whether the counterparty is changed |
| Result re-ranking | Enabled as the hypothetical test starting point | When candidates are numerous, compare ordering before and after re-ranking and check the re-rank score against the reference content |
The optional paths shown in the overall retrieval flow are not executed on every request. Image retrieval requires an image input and the corresponding model capability. Re-ranking requires a model and must be enabled in the settings. A single retrieval run may combine text, image, semantic and full-text paths, and the final reference content is not necessarily ordered by a single vector similarity score.
Three common errors
- Treating a retrieval hit as an authorisation conclusion. The input is source text three, the shipment freeze application with pending review status. The erroneous practice is to output "freeze executed". The check is to compare the application status field and the approver comment field; when the status is pending review and the comment is empty, the conclusion remains to be verified. This is a pending request, not an approval and not an execution.
- Merging the agreement validity period and the freeze status into one judgement. The input is the expiry date in source text one and the application status in source text three. The erroneous practice is to infer that the freeze has been executed from the expiry alone. The check is to verify the agreement fields and the application fields separately, and to record the two facts independently. An expired agreement does not establish that a freeze was executed.
- Treating a similarity score as a permission filter. The input is a minimum relevance of 0.4 together with unrelated distributor records. The erroneous practice is to assume the threshold has excluded unauthorised content. The check is to compare the counterparty identifier in each record against Distributor A in the handover request, exclude records that do not match, and record the exclusions. A relevance score is not a category, entity or permission filter, and retrieval is not authorisation.
How to verify the setup
- Input source text one together with the draft clause "Distributor A's credit agreement has expired". Compare the agreement validity period field with the wording of the draft clause. When the expiry date 28 February 2025 matches the draft clause, that clause is supported. When the materials do not record a renewal, the renewal status is recorded as unsupported. The materials do not state whether a renewal exists, and that gap remains a gap.
- Input source text three together with the draft clause "the shipment freeze has been executed". Compare the application status field and the approver comment field. When the status is pending review and the comment is empty, that clause is assessed as to be verified. It is not assessed as executed and it is not assessed as rejected. A pending request is not an approval, and an unsigned termination is not a completed termination.
- Input the handover request together with the reference array returned by retrieval. Compare the counterparty identifiers in the reference content against Distributor A. When all identifiers match, proceed to the next manual verification step. When another counterparty appears, exclude that item and record it. When the reference array is empty, handle the downstream design using an explicit paper input of an empty array; do not infer that a retrieval rule necessarily returns an empty array. The empty-reference output retains its documented condition: the output path still completes.
Across these checks, the same counterparty, the same period and the same field must be compared. A contradiction is recorded only when one field has mutually exclusive explicit values. When a value is not recorded, or when only evidence of a related stage exists, the result is recorded as unsupported or to be verified. A pending review does not demonstrate a lack of operating permission. Signing, reaching cooperation and contract effectiveness are three distinct facts. The material needed for verification must sit within the accessible target retrieval scope or within an explicit manual input; when it is missing, the gap is recorded.