What the documents look like
Commission settlement for creator partnerships happens after a sale closes. The object of reconciliation is the set of valid promoted orders and the after-sales commission deductions applied to them. Working materials concentrate in the Selected Alliance backend, order after-sales data, the finance system and creator contracts. In a store-service and internal-training setting, multi-turn dialogue has to carry follow-up questions about one settlement across several turns: a staff member may first ask about the commission-setting record for a specific livestream session, then ask about the refund deduction detail for that same session. Materials update weekly and monthly, so a single question cannot cover cross-period reconciliation. Prompts must state the material scope and how to behave when evidence is missing.
Retrieval results expose a similarity value through score. That value describes similarity only. It does not act as a permission filter or an entity-isolation filter, so a high score does not place a record inside the current session's authorization boundary.
Constraints at the multi-turn prompts step
The context a multi-turn exchange actually receives depends on how the session is opened. When a non-empty chatId is passed, the system reads the session from its database and uses only the last element of the messages array as the user question; the remaining messages are ignored. When chatId is not passed, no built-in context is used and the exchange is built entirely from the supplied messages. Prompts must distinguish a new session from a continued session so that ignored history is not treated as already carried in.
Memory turns control how many earlier rounds of chat the answer may reference at most. A value that is too small may prevent earlier information from being used. Reply format follows what the downstream step needs; selecting json_schema also requires the corresponding format requirement to be filled in. Hidden AI output, when enabled, keeps generated content from being shown directly while still passing it on through the AI reply output for later nodes.
Assumed drill sample. Material text: "Creator A June 2025 special-session commission-setting record: category commission rate 5%." "Refund deduction detail for the same session in June: 3 refunded orders, with the corresponding commission offset in the July statement." Draft sentence to be judged: "Creator A's June commission was settled in full at 5%, and the refund commission was not offset." This draft sentence conflicts with the offset fact recorded in the refund deduction detail and is judged unsupported.
Configuration and implementation checks
| Configuration item | Assumed trial starting point | Reason and comparison method |
|---|---|---|
| AI model | Choose a model that supports the needed context length and tool calling | Compare the model's displayed maximum context with the length of a single settlement-material request |
| Memory turns | Starting point set to 4 turns | Reconciliation needs to carry the setting record and the refund detail across two follow-up turns; compare with 2-turn and 6-turn trial runs to see whether follow-ups lose earlier content |
| Reply cap | Enabled, keeping the default value | Compare truncation behavior against the length needed for a complete explanation |
| Temperature | Set lower | Settlement reconciliation requires stability; compare repeated answers to the same question against the default value |
| Reply format | Keep default | No downstream step needs fixed-structure reading; selecting json_schema requires filling in the format requirement |
| Hidden AI output | Keep disabled | The training setting needs to display answers; when disabled, intermediate results pass only through the reply output |
These values are assumed trial starting points for this drill and are not official default values.
Three common errors
- Treating messages other than the last one as already carried into context after a non-empty
chatIdis passed. Check input: attach several history messages under the samechatId. Compare: the question content that actually takes effect against the last element of themessagesarray. Observable result: the answer addresses only the last question. - Treating retrieval
scoreas a permission or entity filter. Check input: a commission-record query that spans stores. Compare: the store each returned entry belongs to against the current session's authorization scope. Observable result: score level has no correspondence with whether an entry belongs to an authorized store, and the identifier must be checked separately. - Treating the display configuration returned by an interactive node as already submitted or approved. Check input: the
interactiveelement insidechoiceswhenstream=false. Compare: that element contains onlytypeandparams, while internal fields such asentryNodeIdsare not returned. Observable result: the exchange waits for a continued session, and no business write has occurred.
How to verify the setup
- Input the assumed sample material text and the draft sentence. Compare the draft's "settled in full, not offset" against the offset fact recorded in the refund deduction detail. Expected observation: the answer states that the sentence lacks support and preserves the evidence gap.
- Input two history messages under the same
chatIdplus one new question. Compare the question that actually takes effect against the last element of the array. Expected observation: the answer addresses only the last question, and the first two are not referenced. - Input an empty result
[]as the retrieval hand-off. Compare how the downstream node is designed to handle an empty list. Expected observation: the downstream follows its empty-input branch and does not fabricate rule-returned content.
These are paper checks and were not actually executed.