What the documents look like
MCN work covers several creators under long-term account compliance, so the materials differ from a single brand session. For store service and internal training, the everyday documents are creator signing contracts, training records, violation handling forms, and account schedules. These materials update daily and weekly. When a schedule changes or a violation status moves, a conclusion reached in one turn can stop applying in the next. Multi-turn prompts therefore have to settle three things: which history each turn carries, what structure the reply follows, and how a missing-evidence state is worded. The prompt must state the material scope, the answering task, and the missing-evidence handling so the model does not write in business facts.
History determines the context boundary. When a chatId is passed, the system takes the session from the database, uses the last item in the messages array as the user question, and ignores the remaining messages. When chatId is absent, context is built entirely from the messages array. The two paths carry different history, so checks must separate a new session from a continued session.
Question optimization applies to continuous knowledge-base dialogue. Checks cover reference resolution and question expansion. Whether a short question qualifies depends on whether that capability is enabled; not every short question is rewritten.
For model and length, the maximum context sets how much content a single request can reference, and the memory turn count controls how many earlier turns are referenced. A long document combined with follow-up questions requires checking the model context and the reply setting separately. Reply format follows downstream consumption. Selecting json_schema requires filling in the matching format requirement and checking the returned structure. The same-vector-model requirement for a retrieval set only constrains knowledge-base identifiers that can be associated together. It does not extend to business-registration sites or other external systems. Semantic recall and relevance scores do not handle permission or entity isolation.
Hypothetical paper exercise
This is a paper exercise. It involves no real customer, real company, or legal identity number.
Material text one (creator signing contract excerpt): "This contract takes effect on the date both parties sign. The cooperation period runs from March 1, 2025 to February 28, 2026. Scheduling follows written notice from the agency."
Material text two (violation handling form excerpt): "Creator A used prohibited wording during the April 10, 2025 livestream. Handling status is pending review."
Material text three (account schedule excerpt): "Creator A, April 15, 2025, store training session, status: scheduled."
Draft sentence to be judged: "Creator A's contract is in effect, the April 10 violation has been fully handled, and the April 15 training session can proceed normally."
Constraints at the multi-turn prompts step
The draft sentence and the three source excerpts sit in one comparison. The contract excerpt supports the effective-contract part. The schedule excerpt supports the scheduled part. The violation form gives the status as pending review, while the draft sentence states the violation has been fully handled. Those are mutually exclusive values for the same field, same subject, and same period, so the draft sentence contradicts the source on that point. The scheduled status only shows the item is on the plan. It does not state whether the violation outcome affects execution. The draft sentence turns a schedule status into an execution conclusion, which the source does not support.
Retrieval and forms are not authorization. A knowledge-base search returns text, and an interactive node form collects values. Business verification, approval, database writes, and external synchronization need their own downstream implementation. The exercise contains no authorization record, so the presence of the materials does not make execution authorized.
The prompt should name the three material types and their dates, state the answering task, and define how missing evidence is worded. The memory turn count, reply limit, temperature, reply format, and multimodal setting are chosen against this sample rather than treated as fixed values.
Configuration and implementation checks
The table lists documented settings or implementation checks tied to this page's documents and primary task. The values are starting points for this hypothetical exercise, not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Memory turn count | 3 | The sample has three material questions, covering contract, violation, and schedule. Compare with 1 to observe whether contract effectiveness information is lost. |
| Reply limit | On, set to medium | The sample must output a judgment and its basis. Compare with a very low value to observe whether the conclusion is truncated. |
| Temperature | Low | The judgment task needs stability. Compare with a higher value to observe whether the same input yields different judgments. |
| Reply format | Default | The sample is a judgment sentence read by staff. Compare with json_schema to observe whether a downstream node needs a fixed structure. |
| Multimodal recognition | Off | The sample materials are text excerpts with no image, audio, or video input. |
| Prompt material scope | State the three material types and dates | The sample contains contract, violation form, and schedule. The prompt must mark the scope and the missing-evidence handling. |
Three common errors
- Treating "pending review" as "fully handled." In the sample, the violation form status is pending review and the draft sentence says fully handled. The two give mutually exclusive values for the same field, which is a contradiction. The check compares the same subject, same period, and same field.
- Treating "scheduled" as "can proceed normally." The schedule status only shows the item is on the plan. It does not state whether the violation outcome affects execution. The draft sentence extends a schedule status into an execution conclusion, which is unsupported.
- Treating retrieval or a form as authorized execution. A knowledge-base search and an interactive node form only return text or collect values. Business verification, approval, database writes, and external synchronization require a separate downstream implementation. The sample holds no authorization record, so the materials alone do not make execution authorized.
How to verify the setup
- Input the three material excerpts and the draft sentence. Compare the draft sentence's field values with the source text. Contract effectiveness and the scheduled status are supported. The violation phrase "fully handled" contradicts the source value "pending review," so the sentence is judged not to hold. If the model outputs that the sentence holds, the prompt or the history carried into the turn has not constrained the comparison.
- Input two consecutive turns under the same chatId. Turn one asks whether Creator A's contract is in effect. Turn two asks whether the April 15 training session can proceed. Compare whether turn two carries forward the turn-one contract conclusion and clearly states the violation-status gap. If turn two gives a proceed conclusion directly, the memory turn count and the missing-evidence handling in the prompt have not taken effect.
- Input an empty retrieval result [] with the question about Creator A's violation handling outcome. Compare the downstream handling design. The expected output is unsupported or pending verification, with no "fully handled" statement. If the output gives a definite conclusion, the empty-result handling has not followed the design.
These are paper checks. No deployment or test has taken place. Missing evidence stays missing. Pending review does not constitute approval, and a scheduled status does not constitute execution authorization.