What the documents look like
JD POP third-party store operations sit around open-platform dispute rules and warranty-period service documents. For the procurement and supplier materials task, the daily documents are the product information sheet, the dispatch time agreement, the dispute handling form, and the warranty-period service record. These materials update daily or per order, and their sources are the JD merchant backend, the JD help center, JD open platform rules, and logistics tracking.
For knowledge base retrieval and recall, the product information sheet and the dispatch time agreement often carry precise fields such as identifiers, model numbers, and time-limit days, so full-text retrieval suits that kind of lookup. The dispute handling form and the warranty-period service record are mostly natural-language narrative, so semantic retrieval or hybrid retrieval suits them better. A retrieval result is only quoted content. It does not constitute authorization or business execution.
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 with different vector models cannot be associated together. The vector model identifier is checked and non-matching bases are excluded; a cross-model retrieval is not run to compare missed recalls. Quoted content is output as an array, and its length can be 0, which means the downstream path still runs even when nothing is found.
The search mode is compared across semantic, full-text, and hybrid retrieval for the test question, and no single mode is presumed to suit all materials. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The quote limit is measured in tokens. Question optimization performs coreference resolution and question expansion, and adds one model call. RRF is the fusion process for hybrid retrieval results, not an on switch. Image retrieval requires an image input and the corresponding model capability, and reranking requires a model and must be enabled by configuration.
Hypothetical paper exercise sample: the following is fictional material, used only for checking.
Material text one (product information sheet excerpt): "Product ID JD-POP-8821, supplier code SUP-045, purchase unit price 12.50 yuan, effective date 2025-03-01, expiry date 2025-08-31."
Material text two (dispatch time agreement excerpt): "For JD-POP-8821 supplied by SUP-045, dispatch within 48 hours after payment is promised, warranty period 12 months, counted from the date of receipt."
Draft sentence to be judged: "SUP-045's supply obligation for JD-POP-8821 continues until 2025-08-31, and the warranty period runs 12 months from the date of receipt."
This draft sentence merges the effective interval of the purchase unit price with the warranty start into one fact. Material one states only the effective and expiry dates of the unit price. Material two states only the dispatch time limit and the warranty period. Neither states a supply obligation term. The judgment requires the same subject, the same period, and the same field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the bases containing the product information sheet and the dispatch time agreement | Check that the vector model identifiers of each base match before associating, and exclude those that do not match |
| Search mode | Hybrid retrieval | Test identifier-type questions and natural-language questions separately, and compare the two hit situations |
| Quote limit | Hypothetical starting point of 2000 tokens | Test with actual materials of different lengths to see whether content is complete, then adjust |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then tune the threshold; it takes effect only under semantic retrieval or reranking |
| Question optimization | Enabled | Compare the original question with the optimized question, and check whether the subject was changed incorrectly |
| Result reranking | ReRank enabled | Confirm the rerank model is available, and compare the ordering before and after reranking |
These values are hypothetical starting points for this exercise, not official default values.
Three common errors
- Treating the expiry date of the purchase unit price as the basis for the end of the supply obligation. Material one states only the effective and expiry dates of the unit price and does not state a supply obligation term; the two are different fields.
- Treating a retrieval hit as authorized execution. Retrieval returns quoted content; business verification, approval, database writes, and external synchronization require separate downstream implementation.
- Associating knowledge bases that use different vector models together. The documentation requires associating knowledge bases with the same vector model; non-matching ones are excluded and the actual observation is recorded.
How to verify the setup
- Input the original text of material one and material two, and compare the draft sentence "the supply obligation continues until 2025-08-31" against the originals. The originals do not state a supply obligation term, so the statement is unsupported and should be recorded as pending verification.
- Input the question "What is the purchase unit price of JD-POP-8821" and compare whether the returned quote contains material one. If the quote contains material one and the unit price field is complete, the expectation is met; if the quote is empty, record the empty result and check whether the data has finished training.
- Input the question "How long is the warranty period for SUP-045" and compare whether the returned quote contains material two. If the quote contains material two and the warranty period field is complete, the expectation is met; if unrelated content is returned, record it and check the minimum relevance and the quote limit.
The above is a paper checking design and does not claim that testing has been performed. Empty-result downstream handling can be checked with an explicit paper input of an empty array against the downstream design; it does not claim that a fictional rule retrieval necessarily returns an empty array.