What the documents look like
Customer-service knowledge base maintenance works on FAQ entries, script templates, product parameter tables and escalation tickets. The output supports store service and internal training. It does not approve refunds. In store service, one product question often splits into several follow-up turns, and a customer-service supervisor rewrites standard answers into reusable dialogue scripts. Materials update weekly or monthly. When a new product parameter table is not synchronized into the script template, store agents answer with old parameters, which produces off-topic replies and escalated complaints. The multi-turn prompts step decides whether a follow-up can carry forward the earlier context, so the material must state its scope, the answering task and how missing evidence is handled.
Constraints at the multi-turn prompts step
Multi-turn carry-over depends on the memory turn setting, which controls how many earlier chat turns the AI references at most. A larger value makes earlier dialogue easier to reference, brings in more content and may affect response speed. In store service, agent and customer follow-ups often span several turns, and a small history count loses earlier conditions.
When continuing a session, a non-empty chatId means the session is taken from the FastGPT database, the last item of the messages array is used as the user question, and the remaining messages are ignored. A new session without chatId builds context entirely from the messages array. The two carry different history, so comparison must separate them. Question optimization needs knowledge-base continuous dialogue to check reference resolution and question expansion, and whether a short question applies cannot be assumed in general.
Reply format follows downstream handling. Selecting json_schema requires filling the corresponding format requirement and verifying the returned structure. A similarity value in retrieval results is a relevance score. It does not perform permission or entity isolation, and retrieval itself is not authorization.
Hypothetical paper exercise sample: a store customer-service supervisor prepares a script template for "insulated bottle capacity and lid accessory." The source material for checking reads: "Insulated bottle model A has a nominal capacity of 500 ml, and its lid accessory model is G-500, applicable to the full model A range. A store lid replacement requires the customer to present the purchase receipt." The draft sentence to be judged reads: "When a customer asks for the model A lid model, reply directly with G-500, and a lid can be replaced without checking the purchase receipt." This draft sentence merges the accessory model with the lid-replacement receipt condition. The receipt requirement belongs to the store lid-replacement step, and the accessory model belongs to the parameter answer.
Configuration and implementation checks
The table lists only settings documented in the official material or implementation checks relevant to this page. The hypothetical starting values are for this paper check and are not official defaults.
| Configuration or check | Relation to this task | Hypothetical starting value and comparison method |
|---|---|---|
| Memory turns | Decides how many earlier turns a follow-up carries | Set to a small number of turns, compare with the common follow-up span in store service, and observe whether earlier conditions are still referenced |
| Reply limit | Limits the length of a single answer | Set to a medium length, compare with the length needed for a complete script-template answer, and observe truncation |
| Temperature | Affects answer stability | Set low, compare with the stability required by knowledge-base question answering, and observe whether answers to the same question drift |
| Reply format | Decides whether the answer needs a fixed structure | Keep the default, compare with whether a downstream node reads fixed fields, and select json_schema only when needed |
| chatId passing | Distinguishes a new session from a continued session | Omit it for a new session and pass a non-empty string for a continued session, then compare the history carried in |
| Retrieval result check | Observes matched entries and scores | Record q, a and score in quoteList, and compare them item by item with the script template source |
Three common errors
- Treating a retrieval relevance score as a permission or entity-isolation mechanism. The wrong behavior is assuming that a high-similarity entry automatically excludes other stores or unauthorized records. The check is to inspect datasetId, collectionId and score in quoteList, and confirm that retrieval returns content while authorization is still decided by a downstream implementation.
- Treating the history of a continued session as a new session. The wrong behavior is building context from all messages during a continued session, ignoring that a non-empty chatId uses only the last item as the user question. The check is to compare the history turns actually carried under the two passing methods.
- Treating the display configuration returned by an interactive node as an executed action. The wrong behavior is assuming that the user has selected or the form has been submitted as soon as the interactive field of userSelect or userInput is received. The check is to confirm that the interactive node returns a display configuration, and that continuing the run requires a new session request with a consistent chatId.
How to verify the setup
- Verification one: for the draft sentence in the hypothetical exercise sample, the input is the source material and the draft sentence, and the comparison is whether the draft sentence merges "G-500 applies to the full model A range" with "a lid replacement requires presenting the purchase receipt" into a replacement without a receipt. The observed result is that the draft sentence differs from the source at two conditions, so the draft sentence is not supported by the source and must be split into a parameter answer and a lid-replacement condition.
- Verification two: for the history carried in a continued session and a new session, the input is the complete messages of the same user question under a request with chatId and a request without chatId, and the comparison is the history turns actually adopted. The observed result is that with chatId only the last item is used as the user question, while without chatId all messages participate in building the context, so the two carry different ranges.
- Verification three: for the correspondence between retrieval results and the script template, the input is the quoteList returned by one paper retrieval, and the comparison is whether q and a match the script template source in subject and field. The observed result is that a matched entry with a different subject or field is recorded as pending verification; if an empty list is returned, the downstream handling design is checked against the explicit paper input of an empty result, without inferring that a fictional rule must return empty.