What the documents look like
Inventory procurement and supplier records in wholesale stall trading draw on market price sheets, stall stock notes, WeChat payment records and logistics consignment notes. The documents handled day to day are purchase sheets, handwritten sales slips, stall lease contracts and payment receipts. A shared trait is fragmentation: one incoming batch may be evidenced by a handwritten slip, a WeChat transfer entry and a consignment note together, and the amount, item name and date across these three materials need not match word for word. The chunking step takes these materials and cuts them into complete semantic units that remain retrievable, while preserving the document identity and version that allow a return to the source text. Handwriting, abbreviations and corrections on handwritten sales slips cause the same field to appear in different forms across materials, so conditions, explanations and the content they constrain need to stay together within a chunk. The supplied update frequency is daily and per order; this describes the rhythm at which materials appear and does not imply batch synchronization or a real-time mechanism.
Constraints at the document chunking step
The basic unit of search is the library; collections serve classification management and do not affect search results. One group of source data can map to multiple vector indexes, and a hit on any one of them recalls that source data group. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, and hybrid retrieval merges the two paths through RRF. Minimum relevance takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. Retrieval is a recall action and does not confirm supplier qualification, payment status or contract validity.
Hypothetical drill sample: the source text has three sentences. Sentence one: "On 12 July, stall A purchased 120 pairs of jeans from supplier B at a unit price of 45 yuan, goods payment 5,400 yuan, agreed for consignment to Chengdong Market." Sentence two: "On 12 July, a WeChat payment record shows stall A received a transfer of 5,400 yuan from buyer C, with the note 'jeans goods payment'." Sentence three: "The stall lease contract shows stall A rents a stall at Chengdong Market, with the term running to 30 June of the following year." Draft sentence to be judged: "Supplier B completed delivery on 12 July and the goods payment has been settled." This draft sentence involves two facts, completed delivery and settled payment; the source text records only a purchase agreement and one same-amount receipt, and states no consignment acceptance, inspection or settlement confirmation.
Configuration and implementation checks
| Configuration item | Assumed trial starting point for this page | Reason for the choice and comparison object |
|---|---|---|
| Search mode | Hybrid retrieval | Materials contain handwritten item names and numbers, so both semantic and keyword coverage are needed; compared with pure semantic retrieval |
| Reference cap | Counted in tokens, starting at a value that can hold one full purchase sheet | A single purchase sheet contains item name, quantity, unit price and amount; compared with a top-k approach |
| Minimum relevance | Starting at a middle value, effective only under semantic retrieval or reranking | Compared with a very low value, to observe whether unrelated fragments are introduced |
| Question optimization | Enabled for multi-turn follow-up questions | Compared with disabled, to observe short-question retrieval effect and added latency |
| Result reranking | Enabled when candidates are numerous | Compared with not enabled, to observe ordering changes |
| Image index | Confirmed against vector model and vision model capability | Compared with no image index, to confirm whether images can be recalled |
Three common errors
- Treating a successful upload as an already-updated retrieval state. Upload and training or index status are two separate stages, and after a material change the actual training and index status need to be checked. Executable check: input the same changed purchase sheet, compare the upload time against whether the retrieved content is the changed version, and observe the phenomenon of returned content still being the old version.
- Treating minimum relevance as a permission filter. Minimum relevance only filters low-relevance results and does not distinguish supplier or stall ownership. Executable check: input a keyword from a purchase sheet belonging to another stall, compare whether the returned results include that stall's content, observe the phenomenon of unrelated records being recalled, record the phenomenon and hand it to manual review.
- Treating a retrieval hit as business confirmation. A hit on a payment record does not mean the goods payment has been settled. Executable check: input "goods payment has been settled" from the draft sentence, compare whether the source text states a settlement confirmation, observe the phenomenon that the source text contains only one same-amount receipt and no settlement statement, and judge it as unsupported.
How to verify the setup
- Verification one: input sentences one and two of the source text, compare whether the retrieved content contains both the purchase agreement and the payment record, and observe whether the two passages are chunked into the same source data group or can be recalled by the same question. The condition is met when the returned content can be traced back to the source text and the amount and date agree.
- Verification two: input the draft sentence "Supplier B completed delivery on 12 July and the goods payment has been settled," compare whether the returned content states consignment acceptance or settlement confirmation, and observe the phenomenon that the returned content contains only the purchase agreement and the same-amount receipt. Under this phenomenon the judgment is unsupported, and the evidence gap is retained.
- Verification three: input a number-type query term, compare the difference between semantic retrieval and full-text retrieval returns, and observe whether full-text retrieval hits the number while semantic retrieval does not. The condition is met when a number-type query returns the corresponding material on the full-text retrieval path; the specific hit situation is recorded from actual observation, with no guarantee of a first-position hit.