What the documents look like
Fruit-store procurement works from purchase orders, acceptance records, pesticide-residue test reports and loss records. The underlying data comes from wholesale-market quotations, supplier offers, store sales and test reports. These materials are produced by the day or by the batch, and their state and validity range determine which statements a draft can carry. Purchase staff, store fresh-produce supervisors and wholesale-market stalls are the typical parties handling them.
Copy drafting for goods and marketing must draw grade, origin and test conclusions from procurement and acceptance materials, then form outward-facing text. A statement is supportable only when the material records the same subject, the same batch and the same field. When a field is absent, the draft sentence stays unsupported or pending verification rather than becoming a confirmed claim. The consequences of getting this wrong include spoilage loss, food-safety complaints and store stockouts, so the boundary between what the material states and what the draft asserts matters directly.
Constraints at the context limits step
The context and token step has several distinct limits. The maximum context is the length of content a model can reference in a single request. The knowledge-base reference cap controls cited content in tokens. The reply cap limits how long a model answer may be. These are separate; the reply cap does not define the context window.
Fruit procurement materials often carry multiple batches and multiple grades. If the reference cap is set too small, the test conclusion or acceptance status needed for a judgment may be trimmed away. If the reply cap is too low, the draft may lose its basis notes. Memory turns affect how far back a continuous follow-up can look. Question optimization adds one model call before retrieval, which usually improves multi-turn retrieval and also adds latency. Minimum relevance is a 0–1 filter threshold that applies only under semantic retrieval or result reranking, and it does not perform permission or entity isolation. Retrieval is not authorization.
Hypothetical paper exercise sample: on 10 June 2025, a store fresh-produce supervisor receives an acceptance record for a batch of "Hainan Feizixiao lychee." The acceptance record states: "200 boxes arrived, 20 boxes sampled, soluble solids measured at 17.2%, pesticide-residue test report number NY-2025-0610-03, conclusion qualified." The supervisor intends to draft: "This store's Feizixiao lychee comes from Hainan, reaches 17.2% sweetness, and passed pesticide-residue testing." A separate handover request reads: "Based on the above acceptance record and test report, confirm whether the sweetness and test-conclusion wording in the draft matches the materials."
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Reference cap | Hypothetically set to 2000 tokens | The sample materials include an acceptance record and a test report, so the conclusion and grade description must fit. Compare with 1000 tokens and observe whether the test conclusion is trimmed. |
| Reply cap | Hypothetically set to 500 tokens | A draft needs to carry its basis notes. Compare with 200 tokens and observe whether the basis notes remain complete. |
| Minimum relevance | Hypothetically set to 0.3 | The sample question points directly at sweetness and the test conclusion. Compare with 0.5 and observe whether the acceptance-record passage is still retrieved. |
| Question optimization | Hypothetically enabled | The handover request contains the reference "the above," so the subject needs completion. Compare with disabled and observe whether the retrieval terms include lychee and the test report. |
| Reply format | Hypothetically left at default | This task does not require a fixed structure. Reassess json_schema only if a downstream step must read fields. |
| Memory turns | Hypothetically a small number of history turns | The sample is a single-turn request, so history has limited effect. Compare with empty history and observe whether the input stays consistent. |
Three common errors
- Treating the reply cap as the full context window. The reply cap limits only model output length; the maximum context governs how much content a single request can reference. Setting the reply cap to 500 tokens and concluding the context is 500 tokens misjudges the space available to the reference cap.
- Using minimum relevance to exclude unrelated businesses or unauthorized records. Minimum relevance is a 0–1 filter threshold that applies only under semantic retrieval or result reranking. It does not perform permission or entity isolation. A relevance score is not a category, entity or permission filter, and retrieval is not authorization.
- Asserting a draft statement is supported when the material does not state it. The sample acceptance record states a sweetness measurement of 17.2% and a qualified pesticide-residue conclusion. If the draft adds "direct supply from origin" and the material does not record a direct-supply relationship, that statement is unsupported and the evidence gap must be recorded. A pending request is not approval, and an offered option is not execution.
How to verify the setup
- Input the sample acceptance-record text and the draft sentence "This store's Feizixiao lychee comes from Hainan, reaches 17.2% sweetness, and passed pesticide-residue testing." Compare the three elements "comes from Hainan," "reaches 17.2% sweetness" and "passed pesticide-residue testing" against what the acceptance record and test report state. The condition is satisfied if all three elements have a matching field with the same subject and batch. If "comes from Hainan" appears only in a supplier offer's origin column while the acceptance record does not state it, that element remains pending verification. Missing evidence stays missing.
- Input the handover request "Based on the above acceptance record and test report, confirm whether the sweetness and test-conclusion wording in the draft matches the materials." Compare the cited content retrieved with question optimization enabled against the content retrieved with it disabled. The condition is satisfied if the enabled setting retrieves the sweetness measurement from the acceptance record and the conclusion from the test report, and the disabled setting lacks the test-report passage. If both settings lack the test report, check the reference cap or whether the data has completed training. Do not assume a scheduled trigger or hidden control.
- Input an empty-result downstream scenario: assume retrieval returns empty cited content, and use the explicit paper input [] to check the downstream design. Compare how the downstream node behaves when it receives empty citations. The condition is satisfied if the downstream node outputs "no relevant material found" and keeps the pending-verification marker. If the downstream node directly generates an affirmative draft, the receiving logic needs adjustment. This check does not claim that a hypothetical rule search always returns []; it records only what is observed.