What the documents look like
Sample inspection for livestream selling happens before a broadcast starts. The question being answered is whether a product can be sold and whether it is suitable for on-camera demonstration. The working materials are sample evaluation forms, third-party inspection reports, supplier qualification files, and prohibited-item lists. When drafting product and marketing copy, the drafter pulls quotable statements out of these materials.
Two features of this document set shape the drafting task. First, materials update weekly or per item, and the same product can carry different conclusions across batches. Copy may only cite the material that matches the current sample, so the batch identifier and report date matter as much as the conclusion text. Second, a single material can be long: an evaluation form plus an inspection report may run well past the length of a short marketing line. The context window of the selected model determines how much of that material a single request can carry, and the knowledge-base reference limit determines how much retrieved content reaches the model. These are separate controls with separate units, and neither one substitutes for the other.
Constraints at the context limits step
The model context window sets the total content length a single request can reference. The reply cap limits only the length of the model's output. Treating the reply cap as the full context window misstates how much source material can be carried, and treating the context window as a reply cap can truncate a drafted paragraph. The knowledge-base reference limit is expressed in tokens, not in a count of items, because chunk lengths in a mixed library vary widely. A long evaluation form can therefore be partly trimmed even when the retrieval itself succeeded.
Conversation memory rounds add earlier turns to the request. The same question asked with empty history and with history produces different inputs, so any judgment about what a drafted sentence is based on must hold the history condition fixed. Problem optimization adds a model call before retrieval and can expand or rewrite the query; it is a separate setting from the reference limit.
A hypothetical drill sample, used only for illustration in this page: the supplier-submitted sample evaluation form reads, "Sample batch number A-2024-11; submission date 5 November 2024; the testing laboratory report conclusion states compliance with the relevant GB requirements; the report is valid until 4 May 2025." The proposed draft sentence reads, "This product has passed national authoritative testing and is valid long term." The comparison to make is between the proposed sentence and the source text on subject, object, status, and validity range. The source text names a batch, a submission date, a conclusion of compliance with relevant GB requirements, and a validity end date. It does not state national authoritative testing and does not state long-term validity. The phrase "valid long term" is unsupported by this source text; the compliance conclusion applies to batch A-2024-11 and the stated validity ends 4 May 2025.
Configuration and implementation checks
The table covers settings and checks relevant to drafting copy from these materials. Starting values are hypothetical for this page only and are not official defaults.
| Configuration item | Hypothetical starting value for this page | Reason and comparison object |
|---|---|---|
| AI model | A model whose configuration popup shows a maximum context large enough for the combined material | Compare against the combined length of the sample evaluation form plus the inspection report |
| Memory rounds | 0 | Compare against a run with history enabled, to keep the draft judgment tied to the current material |
| Reply cap | Left at default | Compare against the length needed to state the basis and the handling action without truncation |
| Reference limit | Set in tokens to cover one full material | Compare against the token length of a single evaluation form to confirm it is not trimmed |
| Reply format | Left at default | Compare against a downstream step that needs fixed fields parsed from the output |
| Minimum relevance | Left at default | Compare against the documented condition that this value applies only under semantic retrieval or rerank |
The minimum relevance value is a filter threshold, not a final ordering rule, and it does not act as a permission or entity filter. Retrieval does not grant authorization to use a record. Whether a retrieved conclusion is permitted for a given product remains a separate business verification step, and any database write or external synchronization is downstream work with its own implementation.
Three common errors
- Treating the reply cap as the full context window. The inputs are the maximum context shown in the model selection area and the reply cap value. The comparison is whether the two are being used interchangeably. The observable result is that a draft's basis is wrongly judged to exceed the available range when only the output length was constrained.
- Setting the reference limit too low so material is trimmed. The inputs are the configured token value and the actual token length of the sample evaluation form. The comparison is whether the configured value covers the whole material. The observable result is a draft sentence citing content that never entered the context.
- Leaving the history condition unfixed. The inputs are two requests for the same question, one with empty history and one with history. The comparison is the difference in input content between the two runs. The observable result is that the judgment about the draft's basis changes with the history setting.
How to verify the setup
- Input the hypothetical sample evaluation form text above and the proposed sentence "This product has passed national authoritative testing and is valid long term." Compare the phrase "valid long term" against the source text's "valid until 4 May 2025," and against the source text's silence on national authoritative testing. Record the phrase as unsupported where the source text does not carry it. The expected outcome is that the unsupported phrase is flagged, not that the sentence is confirmed.
- Input the same question twice, once with empty history and once with history enabled. Compare the two inputs. Confirm that the judgment rests only on the current material and that history introduced no additional basis. The expected outcome is that both runs are judged against the same source text.
- Input a paper record in which the reference result is explicitly empty, written as []. Check whether the downstream design handles an empty result as empty. Do not assume that retrieval always returns an empty value; the point is to verify the downstream handling of a stated empty input, not to predict retrieval behavior.
Where the source text is silent, the correct record is unsupported or pending verification. An unsigned termination is not a completed termination, a pending request is not an approval, and an offered option is not an execution. Missing evidence stays missing.