What the documents look like
Regional dealer stock replenishment runs as a daily rolling order flow, which separates it from seasonal centralized ordering events. The working document set is replenishment requests, inventory turnover tables, outbound notes and payment collection plans. Sources include dealer ERP, the brand ordering system, warehouse stock records and sales receipts. Weekly and monthly update cycles mean several versions of the same material can coexist over time.
In cross-border multilingual work, one replenishment request may exist in both a Chinese version and a translated version, and field names on inventory turnover tables need to be aligned across languages. Retrieval therefore operates over multilingual, multi-version material whose field structures are not uniform. A vector model can carry cross-language semantic understanding, while exact queries for identifiers and model numbers depend more on keyword hits.
Constraints at the retrieval step
The 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 vector model identifier is checked first, non-matching bases are excluded, and the actual observation is recorded. A cross-model retrieval is not run to compare recall differences.
Cited content is output as an array and its length can be 0. Even when nothing is found, the output path still completes. This behaviour can be used to check how a downstream design handles an empty result.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval relies on keyword hits and suits numbers, model codes and proper nouns. Hybrid retrieval uses both recall paths and merges them through RRF. RRF is a merge step for hybrid results and is not described as a standalone switch. ReRank requires a model and must be enabled by configuration. Image retrieval requires the corresponding image input and model capability. The optional paths shown in the overall flow diagram do not mean every request executes all of them.
The minimum relevance value is a filter threshold. It takes effect only in semantic retrieval or when result re-ranking is used, and it is not a final ordering rule. The citation limit is measured in tokens and controls how much content a single search may cite at most. Question optimization adds one model call before the formal retrieval and is used for coreference resolution and question expansion.
Hypothetical exercise sample
Material text supplied for checking (fictional):
Dealer A submitted replenishment request BH-2025-0312 in March 2025, requesting model X200 for a total of 400 units. The status field reads "pending review by the brand channel manager". In the translated version of this request, the item model reads X200, the quantity reads 400, and the status field is translated as "pending review".
Draft sentence to be judged:
Replenishment request BH-2025-0312 has been approved by the brand channel manager, and the 400 units of model X200 will ship this week.
The draft sentence involves the subject (Dealer A), the object (model X200, 400 units), the status (pending review) and the authorization boundary (brand channel manager review). The material text shows the status as pending review, while the draft sentence states approval and gives a shipping time. The two differ in the value of the status field.
Configuration and implementation checks
| Configuration item | Starting value for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the replenishment base that uses the same vector model | Check the vector model identifier of each base, exclude any with a different identifier, and do not run cross-model retrieval |
| Search mode | Run the sample question with hybrid retrieval first | The sample contains model X200 and a natural-language status question, so semantic and full-text recall content are compared |
| Citation limit | Set to 800 tokens as the starting value for this exercise | Test content completeness with materials of different lengths and compare truncation against a 400-token result |
| Minimum relevance | Set to 0.3 as the starting value for this exercise | Label relevant and irrelevant samples first, then compare filtering differences between 0.2 and 0.4 before adjusting |
| Question optimization | Not enabled for this sample, whose subject is already explicit | Compare the original question with the optimized question and check whether the subject was altered |
| Result re-ranking | No re-ranking model is configured, so it stays disabled | ReRank requires a model and must be enabled by configuration; with no model it remains off |
Three common errors
- Treating a pending-review status as approved. The sample material status field reads "pending review by the brand channel manager", while the draft sentence says "approved by the brand channel manager". The input is the sample material and the draft sentence, the comparison is the value of the same status field, and the two values differ. The draft sentence is not supported on that field.
- Treating a retrieval hit as authorized execution. A retrieved replenishment request only means the material was recalled. It does not mean the review passed or that shipping has been arranged. The input is the retrieval result and the draft sentence, the comparison is whether an approval record exists, and the sample contains only a pending-review status, so the approval fact is missing.
- Associating bases built on different vector models. The input is the set of knowledge bases to be associated, and the comparison is their vector model identifiers. Bases with different identifiers are excluded. This error produces no assertable runtime result; only the identifier check and the exclusion action are recorded.
How to verify the setup
- Input the sample material and the draft sentence, and compare the status field values. The material reads "pending review by the brand channel manager" and the draft sentence reads "approved by the brand channel manager". If the cited content returned by retrieval matches the status field of the material, and the draft sentence does not rewrite pending review as approved, the check is satisfied. The expectation is that the status field keeps the pending-review wording. No actual test is claimed.
- Input the sample question "current review status of BH-2025-0312" and compare the recalled content from semantic retrieval and full-text retrieval. If full-text retrieval hits the model code and the request number, semantic retrieval hits the status description, and the hybrid result merged through RRF covers both in its cited content, the check is satisfied. The failure branch is that either path misses the corresponding content, in which case the actual observation is recorded.
- Input a paper case with an empty result, where the cited content is
[], and compare how the downstream design handles the empty array. If the downstream path still completes when the citation length is 0, the check is satisfied. No claim is made that a fictional-rule retrieval necessarily returns[]; the check uses only the explicit paper input to examine the downstream design.