What the documents look like
Channel receivables collection begins after credit sales have been made. The working focus is credit exposure and cash recovery, not matching buyers with sellers. Brand finance staff, sales back-office staff, and dealer owners cooperate around four document types: the receivables aging table, the demand letter, the credit agreement, and the shipment freeze notice. The aging table and the credit agreement together determine when a shipment freeze is triggered. The demand letter records communication progress. The shipment freeze notice connects to inventory procurement and supplier records. Inventory procurement and supplier records include purchase orders, arrival records, and supplier reconciliation data. These materials share dealer entity, contract number, and credit limit fields with the receivables collection materials. Retrieval must compare the same entity, the same period, and the same field group. A retrieval hit is not an authorization to release goods.
Materials update weekly or monthly. The status fields that matter at retrieval include frozen, pending approval, and unfrozen. These labels describe stages in a workflow, and each stage carries a different level of authorization. A pending request is not approval. An unsigned termination is not a completed termination. An offered option is not execution.
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. The model identifier must be checked first, and non-matching bases must be excluded. Cross-model retrieval must not be run to compare recall differences.
Search mode options are semantic retrieval, full-text retrieval, and hybrid retrieval. The choice is made by comparing results against test questions. No single mode is assumed to suit all materials. The citation limit is measured in tokens. Materials of different lengths are used to test whether content remains complete, and the limit is adjusted afterward. The minimum relevance threshold takes effect only in semantic retrieval or when result reranking is used. Related and unrelated samples are labeled before the threshold is tuned. Question optimization handles coreference resolution and question expansion. The original question and the optimized question are compared to check whether the entity has been altered. RRF is a fusion process for hybrid results and is not written as a toggle. Reranking requires a rerank model and is enabled according to configuration. Image retrieval requires image input and corresponding model capability. Retrieval is a recall action and does not perform permission or entity isolation.
Hypothetical paper test case: A brand finance staff member searches the knowledge base for "Dealer A March 2025 credit agreement shipment freeze conditions." Three source sentences are provided for comparison. First: "Dealer A credit agreement number XS-2025-031, credit limit 800,000 yuan, payment term 60 days." Second: "As of March 31, 2025, Dealer A receivables aging table shows overdue amount 120,000 yuan, overdue days 45." Third: "Shipment freeze notice number DJ-2025-0401, status is pending approval." A draft sentence to be judged: "Dealer A credit agreement is effective, shipment freeze has been executed, shipment can be resumed." This draft sentence conflates three separate facts: agreement existence, freeze pending approval, and resumption of shipment. If retrieval returns the three sentences above, they must be compared by the same entity, the same period, and the same field. Agreement number and credit limit belong to the same field group. Overdue amount and overdue days belong to the aging field group. The freeze status "pending approval" and "executed" are mutually exclusive values within the same field. The draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Receivables collection base and inventory procurement base; verify vector model identifiers match | Documentation requires the same vector model; check identifiers one by one and exclude non-matching ones |
| Search mode | Semantic retrieval first, then full-text retrieval for number-type questions | Numbers, notice numbers, and agreement numbers suit full-text retrieval; compare recall differences using the same question |
| Citation limit | Starting point 2000 tokens | Aging tables and credit agreements have long passages; compare content completeness against a 1000-token result |
| Minimum relevance | Starting point 0.4 | Takes effect only in semantic retrieval or reranking; label related and unrelated samples before tuning |
| Question optimization | Enabled for multi-turn follow-up, disabled for single clear questions | Compare original and optimized questions; check whether entity or period was altered |
| Result reranking | Enabled when candidates are many; requires a rerank model | Rerank score is 0-1, filter by score; compare ordering against non-reranked results |
Three common errors
- Treating a retrieval hit as authorization to release goods. Input "Dealer A freeze status," and retrieval returns "Shipment freeze notice number DJ-2025-0401, status is pending approval." Judging "executed" from this is an error. Pending approval is not approval, and retrieval does not carry authorization judgment.
- Associating knowledge bases across different vector models. The two base identifiers differ, yet a joint association is attempted and recall differences are compared. Documentation requires the same vector model. The correct step is to check identifiers, exclude non-matching ones, and record actual observations. Do not fabricate missed recalls or redundant results.
- Conflating agreement existence, freeze execution, and shipment resumption. Input the draft sentence "agreement is effective, freeze has been executed, shipment can be resumed." The materials state only the agreement number and limit, and the freeze as pending approval. They do not state effectiveness confirmation or shipment resumption authorization. An unstated field is unsupported. It is not a negation or a conflict.
How to verify the setup
- Input the three source sentences and the draft sentence from the hypothetical sample. Compare whether the draft sentence exceeds the scope stated in the materials. Expected observation: the draft sentence writes "pending approval" as "executed" and writes the unstated "shipment can be resumed" as a definite conclusion. The judgment is unsupported.
- Input the same question through semantic retrieval and full-text retrieval separately. Compare whether the returned content covers agreement number XS-2025-031 and freeze notice number DJ-2025-0401. Expected observation: full-text retrieval hits the numbers, semantic retrieval hits semantically similar passages, and neither returns an authorization conclusion. This satisfies the condition.
- Input an empty citation array []. Check whether downstream nodes continue. Documentation states the citation length can be 0, and the output path still runs. Expected observation: downstream does not break because of the empty array, and the empty result is not interpreted as "no freeze" or "cleared for release." This satisfies the condition.