What the documents look like
Wholesale stall trading produces records that are more scattered than regional distributor replenishment. Purchase notes, handwritten sales slips, stall lease contracts and payment receipts are issued by different parties at different times, so the amount, quantity and shipping details for one consignment can sit across several sheets. In cross-border work the same document may also exist in a Chinese version and a version in another language, and retrieval needs to cover both wordings. The brief gives a daily and per-order update frequency; this page only indexes documents that already exist, and that frequency does not change how the search node behaves.
The four document types carry different retrieval problems. Handwritten sales slips are short and often contain a product name, a quantity and a unit price. Purchase notes and payment receipts carry amounts and dates. Stall lease contracts are longer and carry a start date, an end date, a monthly rent and a deposit. A multilingual question may use natural language in one language while the indexed chunk uses another, or it may cite a model number or a proper noun that only exact keyword matching will find.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. Before linking, check the vector model identifier and exclude bases that do not match; do not run a cross-model search in order to compare what was missed.
Search mode can be semantic, full-text or hybrid, and no single mode is assumed to suit every document. Semantic search suits natural language questions and cross-language meaning matching. Full-text search suits numbers, model codes and proper nouns. Hybrid search combines the two paths through RRF. The citation limit is measured in tokens, not in a top-k item count, and the minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires image input and matching model capability, and reranking requires a rerank model enabled by configuration. The optional paths shown in the overall retrieval flow do not all execute on every request. The reference output is an array and its length can be zero, so the downstream chain still runs when nothing is found.
Hypothetical paper exercise. A stall owner, Chen, signs a stall lease contract with a buyer. The contract text reads: "The lease term runs from 1 March 2025 to 28 February 2026, monthly rent RMB 8,000, deposit two months." A separate handwritten sales slip reads: "5 March, 300 pairs of jeans shipped, unit price RMB 45, consigned to a named destination." A draft sentence awaiting review reads: "This stall lease contract was terminated on 28 February 2026." The parties are Chen and the buyer, the object is the stall lease contract, the period is the start and end dates above, and the state recorded in the contract is the lease term. The draft sentence states that the contract was terminated, while the contract text records only the end date of the lease term. Expiry of a lease term and termination of a contract are two different facts. The draft sentence is not supported by the supplied wording and remains pending verification.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only one document base that uses the same vector model | Check the vector model identifier first, exclude bases on other models, and do not run a cross-model search for comparison |
| Search mode | Test semantic search first, then test full-text search on the same question | Use one question in both modes and compare the reference content, observing the difference between code-type and natural-language questions |
| Citation limit | Hypothetical starting point of 2000 tokens | Used to test whether long and short purchase notes and lease contract chunks are cited completely, compared against a shorter limit |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then move the threshold; it takes effect only in semantic search or with reranking |
| Question optimization | Off first, tested with complete questions | Compare the original question with the optimized question and check whether the parties were altered |
| Result reranking | Enabled by configuration when a rerank model is available | Reranking returns a score from 0 to 1 used for filtering, compared against results without reranking |
Three common errors
- Treating the lease end date as a termination fact. In the hypothetical sample the contract text records only the start and end dates of the lease term, while the draft sentence states the contract was terminated. The parties, object and period match, but the field value does not, so the sentence is unsupported and the evidence gap must be recorded.
- Treating a retrieval hit as an authorization or permission decision. A retrieved payment receipt only shows that the receipt was recalled within the retrieval scope. It does not show that the payment was settled or that receipt was confirmed. Retrieval is not authorization, and a pending review does not demonstrate that an operation is not permitted.
- Linking knowledge bases across different vector models. The documentation requires knowledge bases that use the same vector model. When a base with a different model identifier appears, exclude it first, and do not then invent missed recalls or redundant results from a combined link.
How to verify the setup
- Input the stall lease contract text from the hypothetical sample together with the draft sentence "This stall lease contract was terminated on 28 February 2026." Compare the parties, object, period and state for sameness. When the contract text does not record a termination action, the draft sentence is classified as pending verification, not as a negation or a conflict.
- Input the handwritten sales slip text "5 March, 300 pairs of jeans shipped, unit price RMB 45" and query "jeans 300 pairs" in full-text mode. Compare whether the recalled content contains that sales slip chunk. Observe whether the reference content array is non-empty. The failure branch is an array length of zero, in which case record that nothing was recalled and check whether the data has finished training.
- Input a query that concerns only a knowledge base on another vector model. Check the vector model identifier of the linked knowledge bases, exclude those that do not match, and do not run a cross-model search. Observe whether reference content is returned only from bases on the same model. When no result is returned, record it as unsupported rather than inferring a missed recall.