What the documents look like
Peak-sale fulfillment planning organizes material around short-term peak capacity and the arrangement of people, vehicles, and space. The main sources are historical peak-sale orders, platform campaign registrations, WMS capacity, and courier pickup capability. Day-to-day documents fall into four groups: peak-sale forecast tables, staff rosters, courier support plans, and exception daily reports. These update per campaign or daily. The product-information and selection question task needs to retrieve the passages tied to a specific product, a specific warehouse, and a specific campaign from those documents, then answer a selection or scheduling question from them. The material is mostly a mix of tables and explanatory text. A single forecast table may hold a product number, a warehouse code, a peak piece volume, and a time-slot arrangement at the same time, and the recall path for coded fields differs from the recall path for natural-language description fields.
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 across different vector models cannot be associated together. Retrieval uses the knowledge base as its smallest search unit; a collection is only for classification management. The search mode can be semantic retrieval, full-text retrieval, or hybrid retrieval, and hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and must be enabled by configuration. Image retrieval requires corresponding image input and model capability. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold and does not carry permission or entity isolation. Question optimization performs coreference resolution and question expansion and adds one model call. Retrieval itself does not constitute authorization. When retrieval returns nothing, the output array length can be 0, and the downstream chain still runs through.
Hypothetical drill sample: a warehousing and distribution team records on paper material "Peak campaign A, warehouse code WH-A, product number SKU-A, forecast peak volume 12,000 pieces, time slot November 1 to November 3, roster owner Group A, courier support plan number EX-A, status submitted for review." The draft sentence to be judged is: "Courier support plan EX-A for SKU-A at WH-A is in effect, and 12,000 pieces can be arranged for dispatch." This draft sentence concerns the plan's effective status and dispatch authorization, and it must be compared against the source text.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases that contain peak-sale forecast tables, staff rosters, courier support plans, and exception daily reports and that share the same vector model identifier | First check each knowledge base's vector model identifier, exclude those with a different identifier, and do not run cross-model retrieval |
| Search mode | First test the same question separately under semantic retrieval and full-text retrieval | Observe coded questions and natural-language questions separately, then decide whether to switch to hybrid retrieval |
| Citation limit | Set the hypothetical starting point to a value that covers one complete passage of a forecast table | Test with actual material of different lengths to see whether content is truncated, and adjust after comparing against the truncation observed |
| Minimum relevance | Set the hypothetical starting point to a value that separates labeled relevant and irrelevant samples | Takes effect only under semantic retrieval or when reranking is enabled; adjust after comparing against the labeled samples |
| Question optimization | Enable first for multi-turn follow-ups and short questions, disable for precise coded queries | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Reranking | Enable by configuration when a rerank model is available | Compare candidate ordering with and without it, and observe whether it improves |
Three common errors
- Forcibly associating knowledge bases that use different vector models. The input is two knowledge bases with different vector model identifiers; compare the identifier field of each knowledge base, and if the identifiers differ they must not be associated together. Exclude those that do not match and record the actual observation; do not infer a runtime result.
- Treating minimum relevance as a permission or entity filter. The input is a retrieval request that contains unrelated business passages; compare the candidates before and after filtering. If unrelated passages are still recalled, the threshold does not carry an isolation role; permission and entity isolation require a separate system implementation.
- Reading a submitted-for-review status as already in effect. The input is the source text of the hypothetical sample above and the draft sentence; compare "status submitted for review" with "in effect." The two are mutually exclusive values on the same field, so the draft sentence is unsupported. Do not add an effective clause to obtain a definite answer.
How to verify the setup
- For the hypothetical sample, enter the question "What is the status of courier support plan EX-A for SKU-A at WH-A?" and compare the retrieved passage with the source text "status submitted for review." If the returned passage contains that status field and has not been rewritten as in effect, the expectation is met. If the return is empty, record the gap and do not judge it as a negative.
- For the draft sentence "Courier support plan EX-A for SKU-A at WH-A is in effect, and 12,000 pieces can be arranged for dispatch," compare the draft sentence with the value on the same field in the source text. If the source text states only "submitted for review," the sentence is unsupported and must be marked for verification. Do not claim it was actually tested.
- For the empty-result branch, enter the explicit paper input [] and compare whether the downstream chain design accepts a citation content length of 0. If the downstream does not handle an empty array, record it as an implementation gap to be filled. Do not claim that a fabricated rule retrieval necessarily returns [].