What the documents look like
Fruit-store procurement runs on four recurring paper materials: purchase orders, acceptance sheets, pesticide-residue test reports, and loss records. A purchase order carries supplier, item, quantity, quoted price, and the stated validity of that quote. An acceptance sheet carries the arrival grade, piece weight, and deducted loss for the delivered batch. A test report carries a batch identifier and a test conclusion. A loss record carries the quantity and cause of spoilage observed at the store. These materials update daily or per batch, and field conventions differ between wholesale stalls, buyers, and store staff.
Compared with convenience-store prepared-food ordering, fruit purchasing leans more on wholesale-market conditions and grade-based acceptance, and less on a headquarters standard meal list. That difference matters at the vector-index step because the same item name can appear with different grades, batches, and quote windows across documents. A retrieval result supplies candidate source text only. It does not carry an acceptance conclusion, a payment approval, or a supplier admission decision.
The knowledge bases joined to a retrieval must use the same vector model. The actual model and its supported input types are checked first. One source data block can be associated with multiple indexes, so index content and source text are checked separately. Index content is built from complete and focused information, and retrieval coverage against index length is weighed against the sample in hand. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion, and cannot be judged from the query side alone. After a model or index change, the same set of questions is compared against the cited source text. A vector score on its own does not support a claim about outcome.
Constraints at the vector indexes step
The primary task here is retrieval over procurement and supplier materials. The retrieval layer returns candidate source text; business verification, approval, database writes, and external synchronization need their own downstream implementation. Semantic recall and relevance scoring do not perform permission filtering or entity isolation. A relevance threshold is not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorized records.
A linked knowledge base must use the same vector model as the others in the same retrieval. Where model identifiers do not match, the mismatched base is excluded before any comparison. Cross-model retrieval is not run in order to compare missed recalls. The requirement that linked bases share a vector model does not extend to industry-conditions websites or other external systems.
Index content and source content are checked as separate objects. A single source data block may map to several indexes, and retrieval can return the block when any one of its indexes is matched. Index length and retrieval coverage trade against each other, so the choice is made against the sample rather than by a fixed rule. Image retrieval paths depend on the model capabilities in use and on whether an index was created at ingestion; an image input alone does not guarantee a match.
The documented retrieval flow lists optional paths, and not every path runs on every request. Image retrieval requires a corresponding image input and the relevant model capability. Rerank requires an available model and is enabled according to the setting in use. A pending handover request is not approval. An offered option is not execution. An unsigned termination is not a completed termination.
Hypothetical drill sample. A store produce supervisor submits a paper handover request. The source materials read: "Purchase order dated 12 March 2025, supplier Dongcheng Wholesale Stall, item Red Fuji apple, grade Grade 1, quantity 200 boxes, quoted price 85 yuan per box, quote valid until 15 March 2025." "Acceptance sheet dated 13 March 2025, same batch Red Fuji apple, received 198 boxes, accepted grade Grade 1, deducted loss 2 boxes." "Pesticide-residue test report dated 13 March 2025, same batch Red Fuji apple, test conclusion Pass." The draft sentence under review reads: "This supplier's quote validity covers the replenishment supply on 20 March." That sentence joins two distinct facts, quote validity and supply period. The source states only that the quote is valid until 15 March 2025; it does not state a supply period. The draft sentence is therefore unsupported, and the evidence gap is recorded as open.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Purchase-order base, acceptance-sheet base, test-report base, loss-record base | The four bases are checked for the same vector model; each model identifier is reviewed and mismatched bases are excluded |
| Search mode | Hybrid search | Item names and grade terms need both semantic and keyword coverage; cited source completeness is compared against semantic search alone |
| Minimum relevance | Hypothetical starting value 0.4 | This is a drill starting point only; cited source hits for the same question set are compared against 0.3 and 0.5 |
| Citation limit | Hypothetical starting value 2000 tokens | This is a drill starting point only; whether cited source covers subject and scope is compared against 1500 and 3000 |
| Question optimization | Enabled | In multi-turn follow-up cases, cited source is compared before and after enabling, checking for an explicit subject and object |
| Result rerank | Enabled where a model is available | Candidate ordering for the same question is compared against results without rerank; no outcome claim is drawn from this |
The values above are hypothetical drill starting points, not official default values. Database-internal implementation details are outside this table.
Three common errors
- Treating quote validity as the supply period. Input is the purchase-order source text above and the draft sentence. The fields compared are "quote valid until 15 March 2025" and "supply period." The observed result is that the source does not state a supply period, so the draft sentence is unsupported.
- Treating a pending handover request as approved execution. Input is the request's status field. The values compared are "pending review" and "approved." The observed result is that pending review does not demonstrate an absence of operating authority, and it does not constitute approval.
- Linking knowledge bases that use different vector models. Input is the model identifiers of the four knowledge bases. The identifiers are compared for consistency. The observed result is that mismatched bases are excluded, and cross-model retrieval is not run to compare missed recalls.
How to verify the setup
- Input the purchase-order source text and the draft sentence from the hypothetical sample. Compare the quote-validity field with the supply-period field. The draft sentence is supported only if the source explicitly states a supply period covering 20 March. The current observation is that it is unsupported.
- Input the acceptance-sheet source text and the pesticide-residue test-report source text from the hypothetical sample. Compare batch, item, and date for the same subject and the same period. The two materials may be cited together only if all three point to the same Red Fuji apple batch. The current observation is that the batch matches.
- Input a paper empty result []. Compare whether the downstream design carries a citation output of length zero. The condition is met only if the downstream chain still runs with an empty result. No claim is made that a hypothetical rule-based retrieval must return [].
Retrieval and forms collect values. Business verification, approval, database writes, and external synchronization require separately connected system implementations. Semantic recall and relevance do not carry permission or entity isolation.