What the documents look like
Channel receivables collection happens after credit sales, so the material centres on credit granting and cash recovery. The main sources are the finance system, bank receipts, distributor orders and sales contracts. The day-to-day documents are the receivables ageing table, the dunning letter, the credit agreement and the shipment freeze notice. In cross-border work the same set of material may exist in both Chinese and English versions, and the ageing table is updated weekly or monthly.
The subject, period and status of these documents differ clearly. An ageing table records the outstanding balance of one distributor by ageing bucket at a given cut-off date. A dunning letter targets a specific contract and debt period. A credit agreement sets the limit and the payment term. A shipment freeze notice shows that a particular shipment sits in a pending review state or has already been executed. Retrieval has to locate the question to a specific document and must not merge same-named files belonging to different distributors into one record.
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. Retrieval only recalls references from content already stored in the knowledge base; it does not perform authorization judgement. A dunning letter being retrieved does not mean the distributor has been granted an extension or that a freeze has taken effect.
Semantic retrieval suits natural-language questions, full-text retrieval suits exact queries such as numbers and contract identifiers, and hybrid retrieval merges multiple result paths through RRF. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold and does not constitute entity or permission isolation. The reference limit is measured in tokens. In a mixed knowledge base the chunk lengths differ widely, so a token-based limit is more stable than a fixed number of items. Question optimization performs coreference resolution and question expansion and adds one model call.
Hypothetical paper exercise sample: one credit agreement in the knowledge base reads in its original text: "The credit recipient under this agreement is Distributor A of the East China region; the credit limit is RMB eight hundred thousand yuan; the payment term is sixty days after shipment; the valid period runs from 1 January 2024 to 31 December 2024." A separate ageing table states: "As of 30 June 2024, the overdue amount for Distributor A is one hundred and twenty thousand yuan." A draft sentence to be assessed reads: "Distributor A has been approved to extend the payment term to ninety days." This draft sentence has no corresponding basis in the available material and remains pending verification.
Configuration and implementation checks
The table below concerns this page's sample and lists settings or check items relevant to this feature and the primary task. The values are hypothetical starting points for this exercise, used for comparison against another value, and are not official default values.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model knowledge base containing the credit agreement and the ageing table | First check the vector model identifier of each base and exclude those whose identifiers differ |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone for recall of exact terms such as contract numbers and amounts |
| Reference limit | Hypothetical starting point of a given number of tokens | Use the tokens needed for one complete credit agreement as the reference, and compare against a smaller value to see whether content is truncated |
| Minimum relevance | Hypothetical starting point of a middle value | First label relevant and irrelevant samples, then compare the filtered results against a higher and a lower value |
| Question optimization | Enabled | Compare whether the subject in the original question and the optimized question stays consistent across multi-turn follow-ups |
| Result reranking | Decided by whether a usable rerank model is available | When a model is available and enabled, compare the candidate ordering before and after reranking |
Three common errors
- Treating a retrieval hit as an authorization fact. The input is the credit agreement and the ageing table above. If the output is "Distributor A has been approved for an extension", the comparison object is the draft sentence against the original text. The original text states only the limit, the payment term and the valid period; it does not state an extension approval. The sentence remains unsupported.
- Associating knowledge bases built on different vector models. The input is two bases whose vector model identifiers differ. Compare the identifiers. If the identifiers do not match and association is still attempted, this is a configuration error. The documents do not describe the runtime result of such an association; record the identifiers and exclude the non-matching base.
- Treating minimum relevance as a permission or entity filter. The input is a query unrelated to Distributor A. Compare whether the returned content includes records of other distributors. Relevance affects only similarity filtering and does not isolate entities. Returning another distributor's content is a retrieval-scope issue and requires manual review.
How to verify the setup
- Using the sample original text as input, search for "Distributor A credit limit payment term" and compare whether the returned references include the original credit agreement text. If the referenced content appears complete and is not truncated, the expectation is met. If the reference is empty, record it as not recalled and check whether the data has finished training.
- Using the draft sentence "Distributor A has been approved to extend the payment term to ninety days" as the object to be assessed, compare whether the original text states an extension approval. The original text does not state this fact, so the assessment is pending verification and must not be judged as approved.
- Using two bases whose vector model identifiers differ as input, compare whether the identifiers match. When the identifiers differ, exclude joint association and record the actual observation. When the identifiers match, check the retrieval results without presupposing missed recall or redundant results.
An empty result can be carried forward using a clearly labelled paper input of an empty list to check the downstream design; this page does not claim that a retrieval over a fictional rule necessarily returns an empty array. Image retrieval and reranking are optional paths and run only when the corresponding input and model capability are present.