What the documents look like
Douyin storefront shelf operations centre on product cards and the store experience score. The people doing this work are Douyin shop merchants, content commerce operators and product card operators. The documents handled day to day are product qualification pages, experience score breakdowns, shipping violation appeals and store announcements. Data comes from the Douyin shop workbench, the Douyin e-commerce learning centre, Selected Alliance data and logistics interfaces, with updates arriving daily and per order. Copy drafting depends on multi-turn questioning about selling points, qualification boundaries and suitable buyer groups. The number of memory rounds, the scope of supplied material and the way missing evidence is handled all affect whether a drafted sentence can be used.
Constraints at the multi-turn prompts step
In multi-turn dialogue the prompt needs to state the material scope, the answering task and how missing evidence is handled, so the model does not write in business facts that were never supplied. When a session continues, the system takes the session from the database by chatId, uses the last item of the messages array as the user question, and ignores the remaining messages. A new session passes no chatId and builds context entirely from messages. Memory rounds control how many earlier rounds of chat are referenced at most; a larger value makes earlier dialogue easier to reference and also carries in more content. Question optimisation is used to check pronoun resolution and question expansion during continuous knowledge base dialogue. Model and length need checking against maximum context and reply cap. Reply format follows downstream consumption; selecting json_schema requires filling in the corresponding format requirement.
Hypothetical paper exercise sample. Material text: "The thermos sold by this store is made of 304 stainless steel, capacity 500 millilitres, applicable temperature range from minus 20 degrees Celsius to 100 degrees Celsius. The product qualification page records an executive standard of GB/T 29606, a test report number of hypothetical number JC-2024-0317, an issue date of 17 March 2024, and a covered model of BWB-500." Draft sentence to be judged: "This thermos uses 304 stainless steel, capacity 500 millilitres, withstands minus 20 degrees Celsius to 100 degrees Celsius, and conforms to the GB/T 29606 standard."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Memory rounds | Hypothetically set to 3 rounds | Copy drafting needs earlier selling points and qualification boundaries; compare context completeness against 1 round and 6 rounds |
| Reply cap | Hypothetically set to 800 tokens | A single copy draft runs roughly 300 to 500 characters; compare truncation against 400 and 1200 |
| Temperature | Hypothetically set to 0.3 | Copy needs to stay close to the material; compare wording drift against 0.1 and 0.8 |
| Reply format | Hypothetically kept as default text | There is currently no downstream structured read; compare downstream consumption cost against json_schema |
| Multimodal recognition | Hypothetically off | The material text is textual; compare against enabling image recognition to see whether unrelated input is introduced |
Three common errors
- The prompt does not state how missing evidence is handled, and the model writes "food grade", which the material does not record, into the draft sentence. To check, input the material text and the draft sentence, compare whether the draft sentence contains wording not recorded in the original, and treat any such wording as unsupported.
- When continuing a session, more than one message is passed by mistake; the system takes only the last one as the user question and ignores the rest. To check, input a request containing two messages, compare what actually enters context against what was expected, and treat a mismatch as redundant input.
- After selecting jsonschema, the corresponding format requirement is left empty, so the returned structure cannot be read downstream. To check, input a configuration with jsonschema selected, compare whether the returned content contains the expected fields, and treat missing fields as an unimplemented format.
How to verify the setup
- Input the hypothetical material text and the draft sentence, compare whether the material, capacity, temperature range and standard number in the draft sentence each have a corresponding statement in the original, and treat full correspondence as satisfying the condition; record any content not recorded in the original as unsupported.
- Input two consecutive follow-up rounds under the same chatId, compare whether the second round answer cites the capacity information confirmed in the first round, and treat citation as satisfying the condition; record a missing citation as context not carried in.
- Input an empty-result downstream scenario, check the downstream design against the paper input
[], compare whether the downstream treats it as having no reference, and treat that handling as satisfying the condition; record any writing in of business facts as improper downstream handling.