What the documents look like
Peak-sale fulfillment scheduling centers on short-term peak capacity and the arrangement of people, vehicles, and space. The product-information and selection Q&A task draws on peak-sale forecast sheets, staff rosters, courier assurance plans, and exception daily reports. These materials update per campaign and daily, so the same product may carry different forecast order volumes, temporary staffing levels, and courier vehicle arrangements across campaigns or dates. Retrieval over this material must cover product numbers, model designations, proper nouns, and natural-language descriptions, while separating the same product's different scheduling states under different campaigns. Historical peak-sale orders, platform campaign registrations, WMS capacity, and courier pickup capacity form the main data sources, and retrieved results need to be checked against the original text blocks corresponding to those sources.
A knowledge base is the smallest search unit; collections only classify data and do not affect search results. A single data block may map to multiple vectors, so one original block can be recalled through any one of its indexes. Indexes and original text are checked separately.
Constraints at the vector indexes step
Knowledge bases associated for retrieval must use the same vector model, so the actual model and its input types are checked first. One original block may link to multiple indexes, and index content is verified against the original text separately. Indexes built from complete, focused information improve precision, while retrieval coverage and index length are weighed against sample behavior. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time, not only on the query side. After any model or index change, the same question set is used to compare cited original text; a vector score alone does not support a performance claim.
Hypothetical drill sample: a warehouse submits a peak-sale forecast sheet whose original text reads "From June 1 to June 3, 2025, the East China warehouse Zone A is expected to ship 12,000 orders per day on average, the temporary staff roster cap is 80 people, and the courier assurance plan sets courier B's daily pickup cap at 15,000 orders." A separate handover request awaits a decision: "Based on the forecast sheet above, confirm whether temporary staff must be increased to 100 on June 2." The actual order volume for June 2, the current roster headcount, and the approval status are not stated in the original text. The retrieval scope is limited to that warehouse's ingested peak-sale forecast sheets, staff rosters, and courier assurance plans, with no external systems or uningested records.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | East China warehouse peak-sale forecast sheets, staff rosters, courier assurance plans | The three bases use the same vector model; model identifier and input type are checked first, and bases with a mismatched model are excluded |
| Search mode | Hybrid search | Product numbers and model designations need full-text search, natural-language questions need semantic search; the same question set compares cited original text returned under each mode |
| Citation cap | 2,000 tokens as the hypothetical trial starting point | This is a paper-drill starting point for comparing whether cited original text covers the forecast sheet's key fields at different caps; it is not an official default |
| Minimum relevance | 0.5 as the hypothetical trial starting point | This is a paper-drill starting point used only when semantic search or reranking is active, to observe whether filtered cited text contains the subject, object, and date |
| Question optimization | Kept off in this hypothetical trial | The pending request already contains the date and object; results without completion are observed first, then compared with results after enabling it |
| Result reranking | Kept off in this hypothetical trial | The ordering of cited original text without reranking is observed first, then compared with reranking scores and text coverage after enabling it |
Three common errors
- Treating a vector similarity score as authorization or an entity filter. Minimum relevance is a filter threshold, not a final ordering rule, and it does not exclude unrelated businesses or unauthorized records. The input is the pending request and the forecast sheet original text; the comparison is whether the retrieved cited original text contains the June 2 order volume and current roster headcount. When the cited text lacks these fields, the vector score does not support a decision to increase staff.
- Associating knowledge bases that use different vector models. The input is two knowledge bases with different model identifiers; the comparison is the model identifier in the node's association configuration. When identifiers differ, that base does not meet the association condition and is excluded, with the actual observation recorded; no cross-model retrieval is run to compare missed recalls.
- Treating image retrieval as a mandatory path on every request. The input is a text-only pending request; the comparison is whether the retrieval chain includes image-description or image-vector recall. When the request contains no image input, the image retrieval path does not execute, and cited original text comes only from the text side.
How to verify the setup
- Input the forecast sheet original text and the pending request from the hypothetical drill sample, and compare whether the retrieved cited original text contains "From June 1 to June 3, 2025," "East China warehouse Zone A," "12,000 orders per day on average," "temporary staff roster cap is 80 people," and "courier B's daily pickup cap at 15,000 orders." When the cited text covers these fields, the text retrieval and index configuration meet the check condition for this drill; when any field is missing, record the gap and check index content and chunking.
- Input the same pending request and run it under hybrid search and semantic search separately, comparing the cited original text and ordering returned under each mode. When hybrid search returns cited text containing the date and quantity fields from the forecast sheet while semantic search may omit precise numbers or designations, record the difference and check the full-text and semantic retrieval configuration.
- Input a request whose original text does not state the June 2 actual order volume or approval status, and compare whether the retrieved cited original text contains the roster approval record for that date. When the cited text does not contain the approval record, the request remains unsupported; record the evidence gap, do not treat a pending review as approval, and do not treat retrieval results as authorization to execute.