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 main data sources are historical peak-sale orders, platform campaign registrations, WMS capacity, and courier pickup capability. Routine documents include the peak-sale forecast table, staffing rota, courier assurance plan, and exception daily report. Update frequency is per campaign and daily. In this scenario, inventory procurement and supplier materials take the form of: the quantity of materials or goods to procure against a forecast order peak for a given campaign, the corresponding supplier quotation and supply commitment, the procurement contact named in the temporary staffing rota, and arrival delays or quantity discrepancies recorded in the exception daily report. Together these materials determine whether warehouse operations and courier pickup can start on schedule during scheduling.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together; the vector model identifier must be checked first and non-matching bases excluded. Retrieval output is an array of references, and its length can be 0, meaning the output path still runs even when no content is found. The minimum relevance score is a filter threshold that takes effect only in semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. The reference limit is measured in tokens and controls the maximum amount of content referenced per search. When question optimization is enabled as configured, it performs coreference resolution and question expansion and adds one model call. Full-text retrieval suits exact queries such as numbers, models, and proper nouns, while semantic retrieval suits natural-language questions. Hybrid retrieval merges multiple result paths through RRF, and Rerank performs a second ordering of candidate texts. Image retrieval requires corresponding image input and model capability, and reranking requires an available model enabled as configured.
Configuration and implementation checks
The following configuration is built around a hypothetical drill sample for this page. Sample content: a peak-sale forecast table states "From 1 November 2025 to 3 November 2025, Warehouse A needs to procure 5,000 packaging boxes; supplier Jia's quotation is valid until 25 October 2025"; the staffing rota states "From 1 November to 3 November, the procurement contact is Yi"; the exception daily report states "On 28 October, supplier Jia notified a 2-day delay in the arrival of packaging boxes." Draft sentence to be assessed: "Supplier Jia committed to complete the supply of 5,000 packaging boxes before 1 November."
| Configuration item | Hypothetical trial starting point for this sample | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the base containing the three materials above | Check each base's vector model identifier first and exclude bases with different models |
| Search mode | Hybrid retrieval | The sample contains exact items such as dates, quantities, and supplier names plus natural-language description; compare semantic and full-text recall paths |
| Reference limit | Hypothetical starting point of 800 tokens | The sample materials are short; compare whether referenced content is complete at 500, 800, and 1,200 |
| Minimum relevance score | Hypothetical starting point of 0.3 | First mark relevant and irrelevant sentences in the sample, then compare whether irrelevant sentences are filtered at 0.2, 0.3, and 0.4 |
| Question optimization | Enabled | This sentence has a complete subject; compare whether the supplier, object, and quantity are preserved with the setting on and off |
| Result reranking | Enabled | When candidates are numerous, compare ordering changes before and after reranking |
Three common errors
- Treating retrieval results as an authorization basis. In the draft sentence "Supplier Jia committed to complete the supply of 5,000 packaging boxes before 1 November," the word "committed" has no corresponding source text in the sample materials. The sample states only that the quotation is valid until 25 October and that a 2-day arrival delay was notified on 28 October. A quotation validity period does not indicate a supply deadline, and an arrival delay notice does not indicate a new supply commitment. Retrieving that supplier's materials does not mean the supplier has made a supply commitment.
- Associating knowledge bases that use different vector models. The sample materials may be distributed across multiple bases; if the vector model identifiers are not checked first and bases with different models are associated directly, the retrieval behavior does not meet the association condition described in the documentation. The correct approach is to check identifiers first, exclude non-matching bases, and record the actual observation.
- Treating the minimum relevance score as a permission or entity filter. The minimum relevance score is a filter threshold that takes effect only in semantic retrieval or result reranking, and it does not exclude unrelated businesses or unauthorized records. If supplier Jia's and supplier Yi's materials are stored in the same base, raising the minimum relevance score will not automatically isolate the two suppliers' data.
How to verify the setup
- Input the three sample source sentences and the draft sentence, and compare the difference between "committed to complete the supply before 1 November" in the draft sentence and "quotation is valid until 25 October 2025" and "2-day arrival delay" in the source. Expected observation: the supply commitment in the draft sentence has no corresponding statement in the source, and is judged unsupported.
- Input the sample draft sentence with question optimization enabled, and compare whether the subject of the optimized question matches the original question. Expected observation: subject information such as supplier Jia, packaging boxes, and 5,000 is not altered in the optimized question. If the subject changes to supplier Yi or another object, record that observation.
- Input the three sample source sentences and run semantic retrieval, full-text retrieval, and hybrid retrieval separately, comparing the recalled content of the three paths. Expected observation: full-text retrieval hits exact items such as "5,000" and "25 October 2025," semantic retrieval hits the natural-language description sentence, and hybrid retrieval, after RRF merging, returns references covering both types of information. If a path does not hit, record that path's input and observation without inferring a cause. For an empty result, use the paper input [] to check whether the downstream path runs.