What the documents look like
Brand omnichannel stock transfers move goods between stores and warehouses. The work does not involve marketing outreach. Data comes from ERP inventory, store POS, e-commerce orders and WMS. Daily documents include transfer requests, inventory snapshots, store shortage lists and transfer logistics notes. Staff service and internal training conversations answer questions such as: what state a given transfer request currently holds, which store a shortage record belongs to, and whether a logistics note has reached receiving. These fields sit in different documents, with updates at daily or weekly frequency. A conversation must carry the subject and period across turns so that the state of one transfer is not applied to the next.
Constraints at the multi-turn prompts step
Multi-turn dialogue depends on memory turns to bring earlier content forward. Too few turns loses the earlier subject; too many increases consumption. The prompt must state the document scope, the answering task and how to handle missing evidence, and must not let the model write in business facts. In continuous knowledge-base conversation, a short question may contain a reference, so reference resolution and question expansion need checking rather than assuming every short question qualifies. Reply format follows downstream consumption; when json_schema is selected, the corresponding format requirements must be filled in and the returned structure checked. Retrieval similarity affects recall ordering only and does not carry permission or entity isolation; retrieval is not authorization.
Hypothetical paper exercise sample: a store shortage list reads "Store ID S-021, item ID P-3307, shortage quantity 12, record date 2024-06-03, status is awaiting replenishment." A draft sentence reads "P-3307 for S-021 has completed transfer from the central warehouse." Checking compares the subject (S-021, P-3307), the status field (awaiting replenishment) and the draft sentence's claim of completed transfer. The two do not agree, and the draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Memory turns | Check the context turns actually carried in | Must cover the shortage list and the follow-up |
| Prompt | State document scope, task and missing-evidence handling | Restrains the draft sentence from adding a completed state |
| Reply format | Choose by downstream consumption; json_schema needs format requirements | Check when a handover request needs fixed structure |
| Model and maximum context | Check the context length the model supports | Check when long documents combine with follow-ups |
| Retrieval similarity | Affects recall ordering only | Not used to exclude unrelated stores or unauthorized records |
| Hide AI output | Content is not shown directly and can pass to later nodes | Check when internal sorting and final reply are separated |
Hypothetical starting value for this exercise: memory turns set to 4, because the sample contains a list and a follow-up, and 4 leaves room before and after; compare with 2 to see whether the subject is lost. This value is not an official default.
Three common errors
- Treating a similarity score as a permission or store filter. Input is the score from one retrieval return; the comparison is whether it is used to exclude unrelated store records; the observable result is that score only orders results and does not isolate them, so unauthorized records can still appear in recall.
- Reading the awaiting-replenishment state as a completed transfer. Input is the sample list's status field and the draft sentence; the comparison is between the two values; the observable result is that the states are mutually exclusive and the draft sentence is not supported.
- Treating an interactive node return as already executed. Input is one userSelect return; the comparison is whether a business action is deemed complete from it; the observable result is that the return is display configuration only, and execution needs a separate system implementation.
How to verify the setup
- Input the sample list text and the draft sentence; compare the subject, the status field and the draft sentence's claim; observe whether the draft sentence agrees with the status field. Disagreement means it is not supported.
- Input the score list from one retrieval return and the store scope; compare whether score is used to exclude unrelated stores; observe whether unauthorized records still appear. Appearance means similarity is not carrying isolation.
- Input an empty result []; compare whether the downstream design treats the empty result as no record; observe whether it is misread as confirmed no stock. Misreading means the downstream logic needs correction.
All three checks address the input, comparison and observation of multi-turn prompts, and no test is claimed as performed. Fields lacking evidence remain unknown; awaiting review is not approval; an interactive return is not completed execution.