What the documents look like
Livestream after-sales work tends to arrive in a burst once a single sales session ends. Refund requests, reshipment orders and order lists enter at the same time, and the customer service team has to explain the host's on-air promises in a consistent way. Copy drafting for this category depends on three materials: the livestream order list, the refund applications and the script templates. These materials decide which facts a promotional sentence may cite.
The same subject, the same session and the same order scope must be used when a drafted sentence is checked against the source. A sentence drawn from one session cannot be validated against materials from another. The documents update after each session and daily, so the verification window follows the session and the order range rather than a single message.
Two features of this category shape the drafting task. First, the burst pattern means a large volume of after-sales records can share one host promise, and a single unsupported sentence can spread across many replies. Second, the promise is often made verbally during the session and recorded only indirectly in the order list or reshipment note, so the drafting step has to compare proposed wording against whatever written record exists and leave gaps visible when the record is incomplete.
Constraints at the context limits step
The context and token step decides how much material the model can read in one request, how much knowledge-base content can be cited, and how many earlier conversation turns are retained. Longer materials are more likely to reach the model's maximum context. The citation limit is controlled by tokens, and content beyond the limit is trimmed, which can leave a drafted sentence without support. A large memory-turn setting pulls in unrelated earlier dialogue; a small one prevents the model from using earlier promises. The reply limit only caps output length and is a separate quantity from the full context window.
Hypothetical paper exercise sample: the source text reads "Order scope for this livestream session: 2025-06-01 20:00 to 23:00, order number range A1001 to A1200. At 21:15 the host stated that the gift for this batch ships together with the main item. Warehouse reshipment note B-07 records that on 3 June a gift was reshipped separately for A1032." The drafted sentence under review reads "All gifts for this session's orders shipped together with the main item." This sentence conflicts with the fact in the source that A1032 received a separate reshipment, so it is unsupported.
The same exercise can be used to check a handover request. A request that asks the drafting step to confirm "all session orders shipped together" without naming the order range or the reshipment note leaves the supporting evidence unidentified. The request stays pending until the order range and the reshipment record are supplied.
Configuration and implementation checks
The table below lists, for the sample above, documented settings or implementation checks that relate to this feature and this drafting task. The values are hypothetical starting points for this exercise, used for comparison with the sample material, and are not official defaults.
| Configuration item | Hypothetical starting point and reason | Check method |
|---|---|---|
| AI model maximum context | Choose a model whose displayed value can hold the three materials; compare against the value shown in the configuration dialog | Send the materials and the drafted sentence together and observe whether any part is truncated |
| Memory turns | Set a small number of turns, since customer service scenarios usually retain only a few earlier turns | Compare the same question with empty history and with history, and note the input difference |
| Reply limit | Keep enough length to express the supporting basis and the handling action | Check whether the reply is cut off before the supporting sentence appears |
| Citation limit | Controlled by tokens; combine single-chunk length with the question sample | Check whether the cited content covers the order number range and the reshipment note |
| Minimum relevance | Takes effect only under semantic search or result reranking | Observe whether low-relevance fragments are filtered; treat this as a relevance threshold, not a permission filter |
| Reply format | Select json_schema and fill in the format requirement when fixed fields are needed | Verify that fields are complete and parsing succeeds, rather than checking text length alone |
The model maximum context is displayed per model, and the configuration dialog shows only the options the current model supports. The minimum relevance value applies only under semantic search or result reranking. The citation limit is expressed in tokens, and the reply limit is a separate output cap. The reply format option json_schema requires the corresponding format requirement to be filled in.
Three common errors
- Treating the reply limit as the full context window. When input material exceeds the model's maximum context, the support for a drafted sentence is trimmed while the output can still look complete. Check: compare the total input length against the maximum context shown in the dialog, and observe whether the supporting sentence appears in full.
- Writing the citation limit as a top-k count or a fixed number of items. The citation limit is controlled by tokens, and chunk lengths vary widely. Check: use the same question with chunks of different single-chunk lengths, and observe whether the cited content is trimmed by token count.
- Treating minimum relevance as a permission or entity filter. This value only filters low-relevance results and does not exclude unrelated business lines or unauthorized records. Check: submit a query that includes an unrelated business line, observe whether the returned content can still include non-target orders, and record the actual observation instead of inferring that the result must be empty.
How to verify the setup
- Enter the hypothetical source text above together with the drafted sentence, and compare the sentence "All gifts for this session's orders shipped together with the main item" against the source fact that A1032 received a separate reshipment. The condition is met if the output identifies the sentence as unsupported or asks for the order scope to be supplied; the condition is not met if the output accepts the sentence directly.
- Enter the same question once with empty history and once with history, and compare the two inputs. The condition is met if the history version brings in the earlier promise and stays within the memory-turn setting; the condition is not met if the history has no effect or exceeds the context.
- When a fixed reply format is used, enter the materials and the drafted sentence, and check whether the returned fields are complete and can be parsed. The condition is met if the fields are complete and parsing succeeds; the condition is not met if fields are missing or parsing fails. For an empty-result handover, use the explicit paper input [] to check the downstream design, and do not claim that a fictional rule search necessarily returns [].
These checks describe expected observations for the hypothetical sample and do not report a completed test, an approval or a deployment. Any field not supplied in the source material remains unknown, and a pending request stays pending until the order range and reshipment record are provided.