What the documents look like
Wholesale stall trading is more fragmented than regional distributor replenishment. It leans on on-site price negotiation and consignment shipping, so records arrive per transaction rather than in tidy batches. The documents this page covers are purchase slips, handwritten sales slips, stall lease contracts, and payment receipts. Market quotations, stall inventory, WeChat payments, and logistics consignment notes are the surrounding data sources. Updates happen daily or per order.
For staff training and storefront service, citation tracing means matching every quoted statement back to the exact source document and stating the document's current status. A training answer that cites a sales slip must reproduce the slip's own wording. A fluent answer does not substitute for source evidence.
The knowledge base search node outputs citations as an array, and the array length can be zero. When nothing is retrieved, the output path still runs through. Before linking knowledge bases, staff must confirm that the selected bases share the same vector model; only then can they be associated together.
Constraints at the citation tracing step
Training material that quotes a purchase slip or sales slip must check the original text line by line. The citation limit controls how many tokens of referenced content are returned. The minimum relevance value takes effect only under semantic retrieval or result reranking. Both are retrieval parameters; neither carries permission or entity isolation. Retrieval is a recall process. Business authorization, approval, and database writes require separate downstream implementation.
The following is a hypothetical paper exercise, used to show the checking method. It does not represent a real record or a completed operation.
Hypothetical exercise sample. Source text one: "On 10 March 2025, Stall A issued a handwritten sales slip to Buyer B for two boxes of dried goods, amount eight hundred yuan, balance unpaid." Source text two: "On 12 March 2025, Stall A and Buyer B signed a stall lease contract, lease term starting 1 April 2025, both signature fields completed." Source text three: "On 15 March 2025, Stall A received a payment receipt from Buyer B, noting settlement of the balance on the 10 March sales slip." The draft sentence to be assessed: "Buyer B has settled all payment, the stall lease contract is effective, and Stall A can arrange consignment shipping." This draft sentence involves three separate facts: settlement of the sales slip balance, effectiveness of the lease contract, and arrangement of consignment shipping. Each must be checked separately.
Configuration and implementation checks
The table below lists documented settings or implementation checks relevant to this page's feature and primary task, tied to the sample above. The numeric values are hypothetical starting points for this exercise, used to explain the comparison method. They are not official default values.
| Configuration or check item | Hypothetical starting point | Comparison method and notes |
|---|---|---|
| Associated knowledge bases | Select the purchase slip base and the sales slip base that share one vector model | Check the vector model identifier first; bases with different vector models cannot be associated together |
| Search mode | Hybrid retrieval | Uses semantic recall and full-text recall together, merged through RRF |
| Citation limit | 2000 tokens as the hypothetical starting point | Controls referenced content length by tokens; compare against the total source text of the sample |
| Minimum relevance | 0.5 as the hypothetical starting point | Takes effect only under semantic retrieval or result reranking; compare against the rerank score |
| Result reranking | ReRank enabled | Reorders candidate text a second time, score range 0 to 1; compare against minimum relevance |
| Citation output | Accept the array output | Array length may be zero; an empty array does not mean the workflow has errored or terminated |
Three common errors
- Treating a retrieval hit as business authorization. Input: the sample draft sentence and the sales slip citation returned by retrieval. Compare the retrieval output against the authorization status field. If the citation shows only the slip content and no settlement or approval status, the balance cannot be judged settled and consignment shipping cannot be judged ready to arrange.
- Treating a pending state as a completed fact. Input: the lease contract source text. Compare the signature field entries against the contract's effectiveness clause. If the source text records only that both signature fields are completed and does not record an effectiveness condition, contract effectiveness remains unsupported and the evidence gap must be recorded.
- Associating knowledge bases with different vector models. Input: the vector model identifiers of the purchase slip base and the sales slip base. Compare whether the two identifiers match. If they differ, exclude that base; do not run a cross-model retrieval to compare missed recall.
How to verify the setup
- Input source text one and the balance settlement portion of the draft sentence. Compare the unpaid balance state on the sales slip against the settlement note on the payment receipt. Observe whether both point to the same party, the same period, and the same field. If they match, balance settlement is supported. If the receipt is missing, record the item as pending verification.
- Input source text two and the contract effectiveness portion of the draft sentence. Compare the signature field entries against the effectiveness clause in the contract source text. Observe whether the source text records an effectiveness condition. If it does not, contract effectiveness remains unsupported and the evidence gap is retained.
- Input source text three and the consignment shipping portion of the draft sentence. Compare the payment receipt against any consignment authorization or commission record. Observe whether a consignment authorization or commission document exists. If only the receipt exists and no consignment commission is present, consignment arrangement is a business action requiring a separate system implementation and cannot be determined by retrieval or a form.
A paper input of an empty array can be used to check downstream handling design. This does not mean a constructed rule returns an empty array in retrieval. Across different vector models, knowledge bases cannot be associated together; check the identifier first, exclude those that do not match, and record what is actually observed. The optional paths in the overall retrieval flow diagram do not all execute on every request. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and must be enabled by configuration.