What the documents look like
Peak-sale fulfillment planning for procurement and supplier records draws on historical campaign orders, platform activity registration, WMS capacity and courier pickup capability. Short-term peak capacity and the arrangement of people, vehicles and space sit at the center of scheduling. The documents update per campaign and per day.
For this page the relevant material covers procurement quantities, supplier delivery dates and authorization status. Citation tracing asks whether a planning decision can be traced back to a specific document, a specific version and a specific signature state. Because the same supplier may show different quotations, delivery dates and authorizations on different dates, a citation must lock the version and the date.
The document set for this category typically contains a peak-sale forecast table, staff shift schedules, a courier support plan and a daily exception report. In procurement and supplier work, the material that matters for tracing includes supplier quotations, written confirmations and seal-use applications. These documents record different facts, and each fact has its own validity window.
Constraints at the citation tracing step
Retrieval returns candidate content. It does not decide whether a supplier is authorized or whether a contract has taken effect. The knowledge base search node outputs a citation array, and the length may be zero. An empty array does not mean the workflow has errored and stopped.
Linked knowledge bases must share the same vector model before they can be linked together. That requirement applies to the vector retrieval scope. It does not extend to business-registration websites or other external systems. The minimum relevance setting is a filter threshold. It does not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a switch that can be turned on. Re-ranking requires an available model and must be enabled in the settings. Image retrieval requires image input and model capability.
The citation cap is measured in tokens. A thinking budget is model reasoning overhead and is unrelated to the citation token cap.
Assume a handover review during peak-sale scheduling. The source material to be checked consists of three sentences. The first sentence: "Supplier A issued a written quotation on 1 March 2025, carton unit price 2.3 yuan, valid until 31 March 2025." The second sentence: "Supplier A confirmed by email on 20 March 2025 that 50,000 cartons can be delivered by 10 April 2025." The third sentence: "The procurement handler submitted a seal-use application on 25 March 2025 to sign an April supply contract with Supplier A; the approval status is pending review." The draft sentence to be judged: "Supplier A has signed the April supply contract, authorizing the purchase of 50,000 cartons at a unit price of 2.3 yuan with a delivery date of 10 April."
The judgment points: a quotation validity period is not a supply period; pending review is not signed; an email delivery confirmation is not contract effectiveness. The three source sentences do not state that the contract has been signed and do not state that authorization has taken effect. The draft sentence's "has signed" and "authorizing" lack support in the source text and should be marked as pending verification.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | How to check |
|---|---|---|
| Linked knowledge bases | Select knowledge bases containing Supplier A's quotation, the email confirmation and the seal-use application | Check the vector model identifier of each base and exclude those with a different vector model |
| Search mode | Hybrid retrieval | Compare recalled content against semantic retrieval and full-text retrieval separately |
| Citation cap | Set to 2000 tokens as the assumed starting point for this exercise | Compare against the length of the three source sentences to confirm they fit |
| Minimum relevance | Set to 0.5 as the assumed starting point for this exercise | Compare against the result of matching the draft sentence with the source text |
| Result re-ranking | Enable ReRank | Confirm the re-rank model is available and compare ordering before and after |
| Question optimization | Off | The sample question is explicit; compare recall with the setting on and off |
The citation cap starting value of 2000 tokens is chosen so the three sample sentences fit within one retrieval. The minimum relevance starting value of 0.5 is chosen as a mid-range threshold for this exercise. Neither value is an official default.
Three common errors
- Treating a quotation validity period as a supply period. In the sample, the first sentence gives a quotation valid until 31 March 2025, while the draft sentence states a delivery date of 10 April. These are different fields and cannot be inferred from each other. Check: input the quotation source text, compare the validity field with the delivery field, and observe whether they are merged into one fact.
- Treating pending review as signed. In the sample, the third sentence gives an approval status of pending review, while the draft sentence states "has signed." Check: input the seal-use application source text, compare the approval status field with the draft sentence wording, and observe whether pending review is upgraded to signed.
- Treating an email delivery confirmation as contract effectiveness. In the sample, the second sentence is an email confirming deliverability and does not state that the contract has taken effect. Check: input the email source text, compare its content with the conditions for contract effectiveness, and observe whether signing or effectiveness clauses are absent.
How to verify the setup
- Input the three sample source sentences and compare the draft sentence against the source text on subject, date and status fields. If the draft sentence contains "has signed" or "authorizing" that the source text does not state, judge it as unsupported and record the evidence gap.
- Input an empty citation array [] and check whether the downstream answer step handles it as designed. An empty array does not mean retrieval failed. Observe whether the downstream step processes zero citations according to its design.
- Input the sample draft sentence as the search query and compare whether the recalled content includes the three source sentences. If the recalled content does not include the source text, check the knowledge base training state, the citation cap and the minimum relevance, and record the actual observation without assuming a hit or an empty result.