What the documents look like
Store counts and loss control work from the gap between book records and physical counts, plus shrinkage. The working objects are count sheets, variance analyses, damage write-off approvals and monitoring access records. Supplier settlement is handled elsewhere. Data arrives from POS inventory, WMS, monitoring systems and the SB/T count standard, and refreshes monthly or quarterly.
For copy drafting in this category, the practical use is internal or external wording around count conclusions: shrinkage cause statements, notices about handling damaged goods, and store loss-prevention reminders. Whether a drafted sentence is supported by the material depends on whether the count record, the variance analysis conclusion and the approval status agree on the same subject, the same period and the same field. The monthly or quarterly rhythm means any quantity or status cited in a draft sits inside a time boundary, and drafting requires checking which period the material covers.
Constraints at the context limits step
Context and token limits constrain three things here: how much content the model can reference in one request, how much retrieved reference content the knowledge base returns, and how much prior conversation is carried in.
The reference cap is controlled by tokens, not by item count. A single count note or variance analysis paragraph can vary widely in length, so trimming by token is more stable than trimming by item count. The minimum relevance value is a filter threshold that applies only under semantic retrieval or result reranking. It does not perform permission or entity isolation, and retrieval itself is not authorization. Model context and reply ceiling are two separate settings; the reply ceiling is not a full context window. When a structured reply format carries copy fields, the check must confirm the fields are complete and parse successfully, not merely that the text is short enough.
Hypothetical exercise sample: a store's March 2025 count record states "snack area closing book inventory 120 units, physical count 108 units, variance 12 units, cause pending investigation." The variance analysis for the same period states "variance in this area 12 units, preliminary judgment: near-expiry write-off not registered in time." The damage write-off approval states "write-off request 12 units, status is pending review." The proposed draft sentence is "In March this store completed write-off of 12 units in the snack area due to near-expiry damage." That sentence records a pending status as a completed write-off, which conflicts with the status stated on the approval.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Reference cap | Hypothetical starting point 2000 tokens | Estimated from the length of count notes and variance analysis paragraphs, compared against the total tokens of the sample material, to confirm key sentences are not trimmed |
| Minimum relevance | Hypothetical starting point 0.5 | Applies only under semantic retrieval or reranking; compared against the score distribution of sample question results, to observe whether the variance analysis passage is filtered out |
| Reply ceiling | Hypothetical starting point 800 tokens | Compared against the target length of the draft sentence plus its basis note, to confirm it can hold the conclusion and the material source |
| Memory turns | Hypothetical starting point 2 turns | Compared against a single-turn drafting scenario, to confirm carried history does not override the current count period |
| Reply format | Hypothetical selection json_schema | Compared against downstream field-reading needs, verifying field completeness and successful parsing |
| Question optimization | Hypothetical off | Compared against the completeness of the sample question, which already contains subject, period and field, so no coreference resolution is needed |
These values are hypothetical trial starting points, not official defaults. RRF is the fusion process for mixed retrieval results and is not a switch setting.
Three common errors
- Writing the "pending review" status from the damage write-off approval as a completed write-off. The approval states pending review; the draft says "completed write-off of 12 units." The two take mutually exclusive values on the same field, which is a contradiction. To check, input the approval text and the draft sentence, compare the status field, and observe whether "pending review" has been rewritten as completed.
- Treating "preliminary judgment: near-expiry write-off not registered in time" from the variance analysis as a confirmed cause. That wording is a preliminary judgment and carries no confirmed conclusion; the draft states it as a settled cause, which is unsupported. To check, compare the analysis wording against the draft's causal phrasing and observe whether a preliminary judgment has been promoted to a settled finding.
- Reading the reference cap as an item count, so a long paragraph is dropped whole. The reference cap is controlled by tokens; if set by item count, a single variance analysis paragraph can be discarded entirely. To check, input the sample material and the reference cap setting, compare whether the returned content includes the variance analysis passage, and observe whether the key sentence is missing.
How to verify the setup
- Input the count record, variance analysis and damage write-off approval from the hypothetical sample, plus the draft sentence "In March this store completed write-off of 12 units in the snack area due to near-expiry damage." Compare "completed write-off" in the draft against "pending review" on the approval, and observe whether the status contradiction is identified and marked for confirmation. The expected result is that the contradiction is flagged; no actual test is claimed.
- Input the sample material with the hypothetical reference cap starting point of 2000 tokens, and compare whether the returned reference content includes the full variance analysis passage. Observe whether that passage is returned complete; if it is trimmed, record the gap. No guarantee is made about first-hit ranking or full recall.
- Input the same question with empty history and with 2 turns of history, and compare whether the two retrieval inputs are identical. Observe whether carried history changes the retrieval terms; if it does, record the difference. For empty-result handling, a clearly written input
[]can be used to check the downstream design; no claim is made that retrieval necessarily returns empty. Libraries across different vector models cannot be associated together; check identifiers only, exclude those that do not match, and record the actual observation.