What the documents look like
On-demand retail store fulfillment processes online orders by the minute, which separates it from periodic stock counting. In the store service and internal training setting, the material to be retrieved covers instant orders, picking lists, out-of-stock reports, and delivery exception tickets. Sources include Douyin instant retail, the Meituan or JD Daojia back office, store inventory, and rider tracking. These records update per order, and on a dense day one store produces multiple out-of-stock reports and delivery exception tickets. Retrieval therefore centers on two questions: what fulfillment state a given order currently occupies, and which operating requirement governs a given exception. The update frequency is minute-level and per order; this page does not change the retrieval chain because of that frequency, so synchronization mechanisms are out of scope.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Retrieval treats the knowledge base as the smallest search unit; a collection only classifies data and does not affect search results. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval fuses both paths through RRF, and Rerank reorders candidate text a second time. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. The reference cap is measured in tokens. Image retrieval requires image input and matching model capability; reranking requires a model and is enabled by settings. Both are optional paths and do not run on every request.
Hypothetical paper exercise sample: the training material of one supermarket home-delivery store reads, "After a picker finds an out-of-stock item, the picker selects out-of-stock on the corresponding line of the picking list and submits an out-of-stock report; the report status is pending review. While pending review, the order must not be marked as picking complete. After review passes, the dispatcher decides replacement or refund." A separate handover request is drafted as: "Once the out-of-stock report is submitted, the store is treated as having completed out-of-stock handling, and the picker may continue to the next order." The objects of judgment are the two facts "report submitted" and "out-of-stock handling completed," within the same order, on the field of report status. The source states that after submission the status is pending review, and pending review does not equal review passed. The draft sentence equates submission with completed handling, which conflicts with the source value on the same field. Authorization boundary: a picker submitting a report is an operating action; replacement or refund is decided by the dispatcher, and the store training material grants the picker no such decision right.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the store fulfillment training base | This base uses the same vector model as the material to be retrieved; check base identifiers one by one, exclude bases with a different vector model, and do not associate across models |
| Search mode | Run one round of semantic retrieval and one round of full-text retrieval | Out-of-stock reports and delivery exception tickets carry numbers and status words, which full-text retrieval tends to hit; natural-language questions use semantic retrieval; compare the reference content of the two rounds |
| Reference cap | Start at 2000 tokens | Picking lists and report material vary in paragraph length; use both long and short real material to check whether reference content is complete, then adjust |
| Minimum relevance | Start at 0.4 | First label several relevant and irrelevant samples, observe whether status description sentences survive filtering, then adjust |
| Question optimization | Enable for multi-turn follow-up; disable for a single clear question | Compare the referent of the original question and the optimized question, and check whether the order subject was changed incorrectly |
| Result reranking | Enable when candidates are many, and check the score | The rerank score range is 0-1; compare the ranking position of the same candidate before and after reranking |
Three common errors
- Treating a retrieval hit as authorization. A retrieved out-of-stock report only shows the material can be cited; it does not show the report passed review, and it does not show the picker holds the right to decide replacement or refund. The input to check is the draft sentence and the training material source text; compare the two status fields "report submitted" and "review passed"; the observed result should be that the two differ.
- Treating minimum relevance as an entity or permission filter. That value takes effect only under semantic retrieval or result reranking and filters low-relevance results; it cannot exclude unrelated stores or unauthorized records. The input to check is the reference content for the same question under different relevance values; compare whether material from another store appears; the observed result should record the actual range of material that appears.
- Treating optional retrieval paths as always executed. Image retrieval needs image input and model capability; reranking needs a model and is enabled by settings; when a condition is absent, that path produces no result. The input to check is a text question with no image; compare whether the output references include an image description or image vector source; the observed result should record the actual enablement state, and a missing condition is recorded as not enabled.
How to verify the setup
- Using the hypothetical training material source text and the draft sentence above, input the draft sentence and run retrieval; compare whether the reference content includes the "pending review" status sentence; the observed result should show that status sentence is cited. If the reference content does not include the status sentence, record it as a retrieval gap and do not conclude the draft sentence holds.
- Using the same question, run semantic retrieval and full-text retrieval separately; the input is "Can an order be marked picking complete while its out-of-stock report is pending review"; compare whether the two rounds of reference content point to the same source paragraph; the observed result should record whether the two rounds agree. When they disagree, keep both rounds and do not merge them into a single conclusion.
- Use empty reference content to check the downstream design. The input is a clearly labelled paper input of an empty array; compare whether the downstream node continues to execute along the reference chain whose length can be 0; the observed result should record whether the downstream path runs through. This check does not assert that a fictional rule retrieval always returns an empty array; it only checks the design for carrying an empty result.