What the documents look like
Wholesale stall trading at a market produces records that are scattered across several hands. Purchase slips, handwritten sales notes, stall lease contracts and payment receipts are each handled by different people, and the product name, quantity and unit price for one purchase may sit on separate sheets. Stall inventory, WeChat payment records and logistics consignment notes exist alongside these paper documents. Updates arrive daily or per order.
These traits shape retrieval directly. Handwriting and abbreviations on a handwritten sales note make word segmentation difficult. Field names on a purchase slip and on a payment receipt may not match. A natural-language query can recall a semantically close record whose subject is a different party. A stall lease contract and a payment receipt belong to different business objects, so the retrieval scope has to be separated by dataset.
The page covers one task: inventory procurement and supplier records, with knowledge-base retrieval and recall. It does not construct business functions that lack a documented basis.
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 on different vector models cannot be associated together. During implementation, check the vector model identifier first, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
A knowledge base consists of library, collection and data. The smallest search unit is the library; a collection is only for classifying and managing data and does not affect search results. One retrieval combines semantic search, full-text search, hybrid search, question optimization and reranking according to the input content and search parameters, then produces cited content after similarity filtering and a citation limit trim.
Cited content is output as an array whose length can be zero. When nothing is found, this output path still runs through. The minimum relevance value takes effect only under semantic search or when result reranking is used. It is a filtering threshold and does not carry permission or entity isolation. Retrieval results do not confirm a stall's identity, the settlement of a payment, or the effective status of a contract.
Hybrid search merges vector and full-text results through the RRF formula. RRF is a fusion process, not a switch that is turned on. Image retrieval requires a corresponding image input and matching model capability; reranking requires a rerank model and must be enabled according to the settings. The optional paths in the overall retrieval diagram do not all execute on every request.
Configuration and implementation checks
Hypothetical drill sample: a stall operator intends to check one purchase. The paper material reads, "On 12 March 2025, Li's stall purchased 20 cases of walnuts from Xinglong Dried Fruit, unit price 180 yuan, payment outstanding, agreed to settle at month end." A separate payment receipt reads, "On 20 March 2025, received 3,600 yuan for walnuts from Li's stall, settled." The draft sentence to be assessed is, "Li's stall settled this walnut payment on 20 March." The retrieval request is "Li's stall walnuts payment settled." The materials do not state which purchase the receipt corresponds to, and manual checking is required.
| Configuration item | Assumed starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only same-vector-model libraries holding purchase slips, sales notes and receipts | Check the vector model identifier first, exclude non-matching libraries, record the actual observation |
| Search mode | Hybrid search | Use "Li's stall walnuts payment" to compare semantic and full-text hits and observe whether identifiers and subjects are recalled |
| Citation limit | Starting point set to cover the full token volume of both materials | Both the receipt and the purchase slip need to enter the citation; compare whether both are complete before and after truncation |
| Minimum relevance | Starting point set to a middle value for this drill | Label relevant and irrelevant samples first, then compare whether irrelevant fragments decrease after a threshold change |
| Question optimization | Enabled for short follow-ups such as "was that one settled" | Compare the original question with the optimized question and check whether subject and object were altered |
| Result reranking | Enabled when candidates are numerous | Compare scores before and after reranking and observe whether the settlement-related sentence ranks earlier |
Three common errors
- Treating a retrieval hit as authorization or status confirmation. The input "Li's stall walnuts payment settled" recalls a receipt fragment, and the payment is then treated as settled. The correspondence between the receipt and the purchase slip is not stated, and the retrieval result does not constitute a fact of settlement.
- Treating minimum relevance as a permission or entity filter. An unrelated query is entered with the expectation that the similarity setting automatically excludes records of other stalls. Minimum relevance only filters low-relevance results and does not isolate business entities.
- Associating libraries on different vector models and then inferring runtime behavior. Check the identifier and exclude non-matching libraries first, record the actual observation, and do not invent missed recalls or redundant returns.
How to verify the setup
- Enter the hypothetical sample text and the retrieval request "Li's stall walnuts payment settled," then compare whether the recalled content contains both the purchase slip fragment and the receipt fragment. If both materials appear in the cited content and the receipt is not directly judged to correspond to the purchase, the condition is met. This is the expectation; it was not actually tested.
- Enter the draft sentence "Li's stall settled this walnut payment on 20 March" and compare it with the two source texts. If the retrieval results cannot supply the correspondence between the receipt and the purchase slip, the sentence is unsupported and requires manual checking; it cannot be judged as settled.
- Enter a clearly unrelated paper query and compare the returned cited content. If an empty array is returned, check whether the downstream design continues on an empty result. If content is returned, record the actual observation. Do not claim that a fictional rule retrieval necessarily returns an empty array.
This page addresses only wholesale stall trading, inventory procurement and supplier records, and knowledge-base retrieval and recall. During operational handover, the authorization status of materials, operations and requests cannot be treated as executed merely because retrieval or a form collected a value.