What the documents look like
On-demand retail fulfillment processes online orders by the minute. Its records come from Douyin instant retail, Meituan or JD.com Daojia back-office systems, store inventory, and rider tracking. Failures produce delivery-late compensation, penalties for falsely marking an order delivered, and customer refunds. Operating reviews and exception retrospectives therefore work across four document types: instant orders, picking lists, out-of-stock reports, and delivery exception tickets. These update per order.
The retrieval question for this task is narrow: for a given order at a given stage, which material supports the recorded state, and what does that material actually say. A knowledge-base search node answers only that retrieval question. It does not confirm that a state was approved, executed, or authorized.
Constraints at the retrieval step
A knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. The referenced content is output as an array, and its length can be zero; when nothing is retrieved, the downstream path still runs.
Semantic retrieval, full-text retrieval, and hybrid retrieval suit different material. Exact queries such as identifiers and model numbers fit full-text or hybrid retrieval better. The reference cap is measured in tokens, not in a guessed item count. The minimum relevance value applies only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds an extra model call before retrieval. Reranking is an optional path that requires a rerank model and a setting that enables it; it does not run on every request. Image retrieval likewise requires image input and matching model capability.
Hybrid retrieval merges multiple result paths through RRF. RRF is the fusion step for combined results, not a toggle. A similarity setting is a relevance threshold; it is not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization.
Hypothetical paper exercise sample. Material one reads: "2025-03-11 14:20, order A-7731 picking completed, one out-of-stock item reported, rider collected at 14:35." Material two reads: "2025-03-11 14:20, order A-7731 picking completed, no out-of-stock report submitted." The draft sentence under review is: "Order A-7731 completed the out-of-stock report and was handed to the rider on 2025-03-11." The two materials share subject, date, and the field "out-of-stock report status," with mutually exclusive values, so that field is contradictory. Rider handover is supported only by material one.
Configuration and implementation checks
| Configuration item | Value or check on this page | Basis and note |
|---|---|---|
| Associated knowledge bases | Associate only one or more bases using the same vector model | Check model identifiers first and exclude any that do not match; do not run cross-model retrieval |
| Search mode | Compare semantic, full-text, and hybrid retrieval separately per test question | No single mode is presumed suitable for all material |
| Reference cap | Hypothetical trial starting value 2000 tokens, compared with 1000 and 4000 for content completeness | This is a hypothetical starting value, not an official default; the unit is tokens |
| Minimum relevance | Label relevant and irrelevant samples first, then compare threshold effects | Applies only under semantic retrieval or result reranking |
| Question optimization | Compare the original question with the optimized question and check whether the subject was altered | Used for coreference resolution and question expansion |
| Result reranking | With a rerank model present and enabled, compare the 0-1 score | Optional path, not executed on every request |
Three common errors
- Forcing association of bases built on different vector models. Input: two base identifiers with different vector models. Compare their model identifiers. Observable result: mismatched identifiers are excluded, and cross-model retrieval is not executed.
- Treating minimum relevance as a permission or entity filter. Input: an unrelated business query. Compare the returned references against the authorized scope. Observable result: the similarity setting does not exclude unauthorized records, and retrieval does not constitute authorization.
- Treating a pending review or an issued quotation as approved or supplied. Input: a handover request stating "order A-7731 out-of-stock report submitted, pending review." Compare the material status field. Observable result: pending review does not show approval, and a quotation validity period does not state a supply deadline.
How to verify the setup
- Search "order A-7731 out-of-stock report status" against the two sample materials. Input: the two source sentences above. Compare the same field value in both materials. Observable result: the two values are mutually exclusive, so the field is judged contradictory, and the evidence gap is recorded.
- Check the draft sentence "Order A-7731 completed the out-of-stock report and was handed to the rider on 2025-03-11" against its sources. Input: the draft sentence and the two materials. Compare the two facts, out-of-stock report and rider handover. Observable result: rider handover is supported; the out-of-stock report is not supported because the field is contradictory.
- Check downstream design against an empty reference array on paper. Input: a clearly labelled hypothetical output of []. Compare whether the downstream node continues under an empty reference. Observable result: the path runs and produces no referenced content. No claim is made that a fictional rule always returns [].
This page addresses only on-demand retail store fulfillment, operating reviews and exception retrospectives, and knowledge-base retrieval and recall. Whether materials, operations, or requests are in a given authorization state at handover cannot be established by retrieval or by a form.