What the documents look like
Short-video affiliate sales relies on long-term content exposure and product-link conversion rather than live explanation. Order fulfilment and after-sales questions therefore often arrive days or weeks after a video is published. Operators work across four document types at once: the short-video script, the product list attached to the cart, the commission plan, and violation feedback. The product list determines which product a given video points to, and the commission plan determines settlement attribution. Both sit apart from order information, which arrives through separate records. The brief gives daily and weekly update frequencies. This page does not depend on batch synchronisation or transaction-triggered mechanisms, so those are not developed here.
For the multi-turn prompts step, the practical consequence is that a conversation must be anchored to the correct video, product, and commission plan before any answer about fulfilment or after-sales status is drafted. The prompt must state the material scope, the answering task, and how to handle missing evidence. It must not let the model fill in business facts that the source documents do not contain.
Constraints at the multi-turn prompts step
Multi-turn conversation depends on how history is supplied. When a non-empty chatId is passed, FastGPT retrieves the conversation from its database and uses the last item in the messages array as the user question; the remaining messages are ignored. When chatId is absent, no FastGPT context is used and the full messages array builds the context. A prompt for after-sales questions should therefore distinguish a new session from a continued one.
Reply length, memory turns, and the model context determine how much material can be referenced. Customer-service and knowledge-base scenarios typically keep a small number of history turns. Retrieval similarity is a ranking signal; it does not perform permission or entity isolation. Retrieval results do not constitute authorisation, and values collected by a form are not proof that a business action was verified or executed.
A hypothetical paper exercise sample, not a real record: the attached product list reads "Video A carries product X, commission plan number C-01, effective 1 March to 31 March." The violation feedback reads "Video A title contains a prohibited term, status is pending rectification." A draft sentence awaiting judgement reads "The order for Video A has been refunded, and commission is settled under C-01." This sentence asserts two business actions, refund and commission settlement. The materials do not state refund status or settlement outcome, so the sentence is unsupported and remains to be verified. Pending rectification does not equal removal.
Configuration and implementation checks
| Configuration item | Setting relevant to this page | How to check |
|---|---|---|
| Memory turns | Controls how many prior turns are carried in; keep a small number for service scenarios | Compare the historyPreview actually carried against the intended turns |
| Reply limit | Controls the length of a single reply | Compare whether the answer is truncated |
| Temperature | Lower values suit service and rule-based scenarios | Observe answer stability across repeated inputs |
| Reply format | Choose json_schema with a format requirement only when a downstream step reads fixed structure | Check the returned structure against the requirement |
| Stop sequence | Use only when output must end at a fixed marker | Check whether output stops at the marker |
| Multimodal recognition | Depends on model capability; image recognition requires an image input | Check model capability against the input type |
Hypothetical starting value for this exercise: memory turns set to 2, because the sample involves one follow-up question; comparing against 0 turns shows whether prior text is carried in. This value is a starting point for this exercise, not an official default.
Three common errors
- Treating pending rectification as removal, or treating pending review as absence of operating permission. In the sample the violation feedback status is pending rectification, and the material does not state removal, so the correct judgement is to be verified.
- Treating retrieval similarity as a permission filter, assuming a high score excludes unrelated or unauthorised records. Similarity affects ordering only and does not perform entity or permission isolation.
- Treating the refund and settlement in the draft sentence as having occurred. The material does not state refund status or settlement outcome, so the sentence is unsupported.
How to verify the setup
- Input the sample source text and the draft sentence, then compare whether the draft claims the refund is complete and the commission settled. Expected: the answer states that the material does not record these two items and preserves the evidence gap.
- Input "Video A title contains a prohibited term, status is pending rectification," then compare whether the answer says the video has been removed. Expected: the answer states that pending rectification does not equal removal.
- Construct a paper input of an empty retrieval result, then compare whether the downstream node handles the empty result as designed. Expected: the downstream node processes the empty result without asserting that a fictional rule search must return nothing. All three are paper checks and were not executed.