What the documents look like
Site selection and store opening work happens before a store begins trading, and its core is lease negotiation and construction delivery. The working materials for one site are a site evaluation form, a lease letter of intent, a renovation budget, and an opening schedule. Supporting inputs include trade-area footfall, map POI data, POS history by store type, and landlord-provided documents. These materials are updated irregularly and once per store, and no single template governs them, so the same field can be written differently across documents.
Copy drafting for goods and marketing draws only on the checkable facts in those materials: area, rent, rent-free period and validity period in the lease letter of intent; line-item amounts in the renovation budget; milestone dates in the opening schedule. A draft sentence may cite only what the materials state. Where a material is silent, the point stays pending verification and is not filled in by the model.
Constraints at the multi-turn prompts step
The value of multi-turn dialogue here is sustained questioning of different fields in the same document set. The prompt states the material scope, the answering task, and how to handle missing evidence. History determines what earlier turns the model can see; a small memory turn count drops earlier material, and a large one pulls in unrelated content. Question optimization addresses reference resolution in continuous knowledge-base dialogue, and whether it applies to short questions is checked against the actual session.
Reply format follows what consumes the output. When a downstream node reads a fixed structure, select json_schema and supply the corresponding format requirement; ordinary copy drafting keeps the default. Model and length are checked against maximum context and reply cap, since long documents combined with follow-up questions readily hit truncation.
Hypothetical paper exercise sample: the source text reads "Party A intends to lease Unit 3-12, third floor, Mall A, to Party B; gross floor area 210 square metres; first-year rent 42,000 yuan per month; rent-free period 60 days; the letter of intent is valid until 30 June 2025." The draft sentence to be judged is "This store is on the third floor of Mall A, with an area of 210 square metres and first-year rent of 42,000 yuan per month, and enjoys two months rent-free before opening." Check points: the area and rent figures match the source; whether "two months rent-free" equals the source's "60 days" requires calendar verification; "before opening" has no counterpart in the source and is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Memory turns | 4 turns | Fields are scattered across several follow-ups; compare with 2 turns to see whether earlier material is lost |
| Reply cap | Enabled, set to 800 | Draft sentences are shorter than full proposals; compare with 2000 to see whether truncation appears |
| Temperature | 0.3 | Fact citation needs stability; compare with 0.8 on whether the same question drifts |
| Reply format | Default | No downstream node reads a fixed structure; compare with json_schema for downstream consumption |
| Multimodal recognition | Image enabled according to model capability | Used when materials contain screenshots; compare with off to see whether image fields are missed |
| Thinking configuration | Default | Field checking needs no long reasoning; compare with standard thinking on response time |
Three common errors
- Treating wording absent from the materials as supported. The phrase "before opening" has no corresponding field in the source, yet it is handled as supported, producing a time qualifier with no basis. Check: input the hypothetical source and the draft sentence, compare whether each qualifier has a source location, and mark any without one as pending verification.
- Writing "60 days" directly as "two months". Calendar months differ in length, so the two are not equivalent. Check: input the source rent-free period and the draft sentence, convert by specific start and end dates, and compare whether the day counts agree.
- Treating the letter of intent's validity period as a signed contract. Validity until 30 June 2025 states only the time range of the letter itself and does not indicate that a contract took effect or that authorization was granted. Check: input the source validity field and the draft sentence, compare whether wording such as "signed" or "in effect" appears, and mark it unsupported if it does.
How to verify the setup
- Input the hypothetical source and the draft sentence, compare whether the area and rent figures match the source word for word, and confirm the output contains no qualifier absent from the source. Expected: figures agree, and unsourced qualifiers are flagged.
- Input the source rent-free period "60 days" and the draft phrase "two months rent-free", convert to day counts by start and end dates, and compare whether the two are equal. Expected: when day counts differ, a rewrite is indicated.
- Input the source validity field and the draft sentence, and compare whether signing or effectiveness wording appears. Expected: if it appears, mark it pending verification rather than as signed.
These checks rest on a paper hypothetical sample and no actual call was executed. When retrieval returns nothing, an explicit paper input of an empty array can be used to check downstream handling design; this does not mean a fabricated-rule query necessarily returns empty. A similarity score does not perform permission or entity isolation, and libraries across different vector models cannot be jointly associated, so identifiers are checked first and non-conforming items excluded.