What the documents look like
Wholesale stall trading is more fragmented than regional dealer replenishment, and it relies on on-site bargaining and consignment shipping. The data sources named for this category are wholesale market quotations, stall inventory, WeChat payment records, and logistics consignment notes. The documents handled day to day are purchase receipts, handwritten sales slips, stall lease contracts, and payment receipts. Updates are described as daily and per order.
For store service and internal training, this matters because the material is uneven. Handwritten slips vary in handwriting and in the short names used for goods. Purchase receipts and sales slips must be checked field by field for product name, quantity, and amount. A stall lease contract and a payment receipt describe different facts, so a retrieval result can only offer candidate content for a person to review. The retrieval step does not decide whether money was collected, whether goods were shipped, or whether a contract took effect.
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. Knowledge bases with different vector models cannot be linked together. The check here is to confirm the vector model identifier and exclude those that do not match; a cross-model retrieval is not run to compare missed recalls.
The output of the node is reference content in array format, and its length can be 0. When nothing is found, the output path still runs through. Semantic retrieval, full-text retrieval, and hybrid retrieval are compared using the same test question; no single mode is assumed to suit all of this category's material. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. RRF is the process that merges hybrid retrieval results, and it is not a switch. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and must be enabled according to the settings.
Hypothetical paper exercise sample: the source material reads, "Stall A issued a handwritten sales slip to Buyer B on March 1, stating product potato, quantity 30 pieces, amount 4,500 yuan, payment not collected; on the same day the purchase receipt stated the same product, quantity 30 pieces, purchase price 3,600 yuan." The draft sentence to be judged is, "Buyer B settled the 4,500 yuan payment on March 1." The subject to check is Buyer B, the object is the 4,500 yuan payment, the state is not collected, and the valid scope is that sales slip on March 1. The draft sentence shares the same subject, the same period, and the same field as the source, and its value is mutually exclusive with the source, so it is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the stall document base that uses the same vector model | Confirm the vector model identifier first, exclude bases with a different model, and do not run a cross-model retrieval |
| Search mode | Test semantic retrieval first, then full-text retrieval | Use the same test question to compare the candidate content returned by the two modes |
| Reference limit | Hypothetical starting value 2,000 tokens | Use purchase receipts and sales slips of different lengths to test whether content stays complete, comparing against 1,000 and 3,000 |
| Minimum relevance | Hypothetical starting value 0.4 | Label relevant and irrelevant samples first, then compare the filtered results against 0.3 and 0.5 |
| Question optimization | Turn it off first, then turn it on for comparison | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Compare after a model is available and the setting is enabled | Compare candidate ordering before and after reranking, and confirm whether the question is complete |
Three common errors
- Treating the minimum relevance value as a permission or entity filter. This value is a filtering threshold; it does not exclude unrelated businesses or unauthorized records, and retrieval does not constitute authorization.
- Writing RRF as an on/off switch. RRF is the fusion process for hybrid retrieval results and has no separate switch.
- Treating a pending review or a quotation validity period as an approval or a supply deadline. A pending review does not show that an operation permission is absent, and a quotation validity period does not indicate a supply deadline.
How to verify the setup
- Enter the source text and the draft sentence from the hypothetical sample above, and compare subject, object, state, and valid scope. When the draft sentence and the source hold mutually exclusive values in the same field, judge it as not supported; do not judge the payment as settled.
- Enter one stall document question and compare the candidate content from semantic retrieval with that from full-text retrieval. Observe whether the same set of documents is hit. The failure branch is an empty candidate set or a missing field.
- Enter a clearly stated paper input of an empty array [] and check whether the downstream design accepts an empty reference. Do not claim that a particular rule's retrieval necessarily returns [].