What the documents look like
Peak-sale warehouse and delivery scheduling centers on short-term peak capacity and the arrangement of people, vehicles and space. The working materials named in the brief are the peak-sale forecast table, staff scheduling, the courier assurance plan and the exception daily report. These materials update per campaign and per day. For copy drafting, a sentence that makes a promise to buyers has to line up item by item with the scheduling material for the same campaign.
The drafting task here is product and marketing copy. Citation tracing at this step means each outward-facing sentence can be matched to a source sentence in the designated material where the subject, object, state and effective scope agree. Anything the material does not state is unsupported; the copy cannot fill the gap on its own.
Two features of this category shape the work. First, the material is campaign-specific and date-bounded, so a promise such as a pickup deadline belongs to one campaign and one date range. Second, the material is a forecast and assurance record rather than a completed-operation record, so a stated cutoff time is a plan, not evidence that every order met it.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. Even when nothing is retrieved, the output path still runs. Once that citation array reaches the answering step, the answer has to be checked line by line against the cited source text.
Retrieval is a recall process. It does not perform permission or entity isolation. The minimum relevance value is a filter threshold, not a final ordering rule. RRF is a fusion step for mixed results, not a switch that can be turned on. Reranking requires a rerank model enabled in settings. Image retrieval requires the corresponding image input and model capability. The optional paths shown in the overall flow diagram do not all execute on every request.
Recency checking uses the current version and a historical version of the document as separate sources. Whether a citation is adopted still depends on review of the effective version.
Hypothetical drill sample
Source text for checking (peak-sale forecast table, campaign X, date range assumed to be the 1st to the 3rd of a month): 1) "The forecast peak order volume for this campaign is assumed value A, and temporary staff are scheduled at assumed value B." 2) "The courier assurance plan for this campaign sets the pickup cutoff time at assumed time C." 3) "The exception daily report for this campaign records an assumed value D of late-shipment orders."
One draft sentence to be judged: "During this campaign, all orders can complete pickup before assumed time C."
The subject (this campaign), object (order pickup), state (completion time) and effective scope (during this campaign) needed for the judgment appear in the material only as a pickup cutoff agreement. There is no record that all orders completed pickup. This draft sentence is unsupported and needs verification or rewriting.
Configuration and implementation checks
| Configuration item | Assumed starting point for this trial | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the knowledge base that records this campaign's scheduling material | Check the selected knowledge base and material version, and confirm the vector model identifiers match before they can be linked together |
| Search mode | Hybrid search | Number and time expressions in the copy suit full-text recall, while semantically similar wording suits vector recall; the two are fused through RRF |
| Citation limit | Assumed starting point of a token amount that covers a complete source paragraph of this campaign's material | Compare against the paragraph length of the source text to confirm the citation can contain the subject and effective scope |
| Minimum relevance | Assumed starting point of a medium threshold | Compare against the relevance scores of recalled results and observe whether the source sentence survives filtering |
| Reranking | Enable the rerank model in settings | Compare against the 0-1 score after reranking to confirm the ordering change of candidate text |
| Question optimization | Decide whether to enable based on actual effect | Compare retrieval results before and after enabling, and observe whether coreference removal improves retrieval |
Three common errors
- Treating retrieval recall as an authorization basis. The input is one retrieval request; the comparison is between the citation array and the source text; the observable result is whether the cited content contains the subject, object, state and effective scope. A high retrieval score does not mean the item is permitted for outward-facing copy.
- Treating unstated content as confirmed. The input is the draft sentence and the corresponding source text; the comparison is between the promise in the draft sentence and the record in the material; the observable result is whether the material states that promise. When the material does not state it, the sentence is unsupported and needs verification or rewriting.
- Treating optional paths as always executed. The input is one retrieval request; the comparison is against the image input, model capability and settings actually present for this request; the observable result is which retrieval paths actually participated. Image retrieval and reranking execute only when their conditions are met.
How to verify the setup
- Enter the source text and the draft sentence from the hypothetical drill sample. Compare the promise in the draft sentence against the record in the source text and observe whether the material states that all orders complete pickup before assumed time C. When the material does not state it, the draft sentence is unsupported and needs verification or rewriting.
- Enter one retrieval request. Compare the citation array against the source text and observe whether the cited content contains this campaign, order pickup, completion time and effective scope. When the citation array length is zero, the output path still runs, so check how the downstream answering step handles an empty array. An empty array can be checked on paper as an explicit input of [] against the downstream design; this does not assert that a hypothetical rule-based retrieval always returns [].
- Enter the current version and the historical version of the material as separate retrievals. Compare the two citation sets against their respective source versions and observe whether each citation comes from the matching version. Whether a citation is adopted still depends on review of the effective version, and version differences are recorded as evidence gaps.
Knowledge bases with different vector models cannot be linked together. Check the identifiers first and exclude those that do not match; do not run a cross-model retrieval to compare missed recall. Retrieval checks record actual observations and do not guarantee a first-hit match, full recall or an empty result for unrelated queries.