What the documents look like
Regional dealer stock replenishment in the wholesale and dealer channel runs as a daily rolling order flow. Replenishment requests are raised on a weekly or monthly cycle, which separates this work from seasonal centralized ordering events. The main data sources are the dealer ERP, the brand ordering system, warehouse stock records, and sales collections. The documents handled day to day are replenishment requests, inventory turnover sheets, outbound delivery notes, and collection plans. This page addresses only knowledge-base retrieval and recall for the inventory procurement and supplier records task. It does not extend to ordering fairs or terminal retail scenarios.
Replenishment requests and inventory turnover sheets carry fields such as model, quantity, and turnover days. Supplier records carry fields such as supply lead time, settlement method, and authorization status. These materials mix natural-language description with identifier-type information. Semantic retrieval suits questions such as whether a given supplier can cover replenishment for this region. Full-text retrieval suits exact hits on a specific model number or document number. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the vector model identifier must be checked first and non-matching bases excluded. The retrieval output is an array of cited content whose length can be 0, and the downstream branch still executes when the result is empty. Retrieval itself does not perform authorization judgment. Whether a supplier is within a valid cooperation period, and whether a replenishment request has been approved, both require manual checking of the source material or implementation in a downstream system.
Constraints at the retrieval step
The lowest relevance score is a filter threshold that takes effect only under semantic retrieval or when result reranking is used. It is not a permission or entity-isolation mechanism. A relevance score does not identify a region, a supplier, or an authorization state. Retrieval is not authorization. When unrelated records appear in the cited content, they are excluded by checking the source text, not by relying on the threshold to isolate them automatically.
The search mode determines which recall paths run. Semantic retrieval depends on vector distance and handles natural-language questions about supplier coverage. Full-text retrieval depends on keyword hits and handles model numbers, document numbers, and proper nouns. Hybrid retrieval uses both and merges the two result sets through RRF. Reranking applies a second ordering pass to candidate text and returns a 0-1 score. RRF is a fusion process for combining multiple recall paths; it is not a switch that is turned on. The optional paths in the overall retrieval flow do not all execute on every request. Image retrieval requires a corresponding image input and the relevant model capability. Reranking requires a rerank model and must be enabled in the settings.
Question optimization performs coreference resolution and question expansion before retrieval. It adds a model call and therefore increases overall latency. It can be compared against the original question to check whether the subject still points to the same supplier. The citation limit is measured in tokens. Because chunk lengths differ widely across a mixed knowledge base, the limit is expressed in tokens rather than top k.
Configuration and implementation checks
The configuration below centers on one hypothetical drill sample. Hypothetical drill sample: a regional dealer sales back-office clerk prepares a replenishment request on 10 March 2025. The request source text reads: "East China regional dealer A submitted a replenishment request on 10 March 2025, covering model X200 for a total of 300 units, target warehouse is the regional central warehouse, request status is pending review by the brand channel manager." A supplier record is also held. Its source text reads: "The supply agreement for supplier B is valid from 1 January 2025 to 31 December 2025, covers the East China region, and the settlement method is monthly settlement at 30 days." The draft sentence to be judged is: "Supplier B has confirmed acceptance of this X200 replenishment and has arranged shipment." Neither "confirmed acceptance" nor "arranged shipment" has a corresponding status in the source text.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the inventory procurement and supplier records base | Check each base's vector model identifier first, exclude non-matching bases, then run retrieval |
| Search mode | Hybrid of semantic retrieval and full-text retrieval | Test the full-text hit with model X200, test the semantic hit with a natural-language question, compare the two result sets |
| Citation limit | Hypothetical starting point of 2000 tokens | Test with the replenishment request and the supplier record, which differ in length, observe whether cited content is truncated, then adjust |
| Lowest relevance score | Hypothetical starting point of 0.4 | Manually label relevant and irrelevant samples first, compare whether the supplier record source text is retained after filtering |
| Question optimization | Enabled | For a short question such as whether that supplier can cover the region, compare whether the subject still points to supplier B before and after optimization |
| Result reranking | Enabled as needed | When candidates are numerous, compare the ranking position of the supplier record source text before and after reranking |
Three common errors
- Treating a retrieval hit as authorization confirmation. The input is the draft sentence "Supplier B has confirmed acceptance of this X200 replenishment and has arranged shipment." The comparison object is the supplier record source text, which states only the agreement validity period and the covered region. It states no acceptance confirmation and no shipment arrangement. Treating the retrieved supplier record as confirmation of acceptance mistakes a retrieval result for business authorization. The correct action is to record that this status is absent.
- Treating the lowest relevance score as a permission filter. The input is "East China regional dealer A's replenishment request." The comparison object is the cited content returned by retrieval, observed for the appearance of records from other regions or unauthorized suppliers. The lowest relevance score filters by score only and does not recognize region or authorization status. When unrelated records appear, they are excluded by checking the source text; the threshold cannot be relied on to isolate them automatically.
- Forcing an association across different vector models. The input is two knowledge base identifiers. The comparison object is each base's vector model identifier. If the two are not consistent, they cannot be associated together. The non-matching base is excluded and the actual observation is recorded. No inference is drawn that a forced association would produce missed recall or redundant results.
How to verify the setup
- For the draft sentence "Supplier B has confirmed acceptance of this X200 replenishment and has arranged shipment," input this sentence together with the supplier record source text. Compare whether the source text contains an acceptance confirmation field or a shipment arrangement field. The condition is met when retrieval returns the supplier record source text, manual checking confirms the source text does not contain those statuses, and the draft sentence is marked as unsupported.
- For the replenishment request source text "East China regional dealer A submitted a replenishment request on 10 March 2025, covering model X200 for a total of 300 units," input model X200 for full-text retrieval and input "East China replenishment request review status" for semantic retrieval. Compare the cited content from both paths. The condition is met when both paths return the request source text, the cited content fully retains the request status as pending review, and no approved wording appears.
- For the empty-result branch, input a paper-constructed cited content array of []. Compare the downstream node design. The condition is met when the downstream chain continues to execute with the empty array, does not halt on the empty result, and no claim is made that a fabricated rule would necessarily return an empty array from retrieval.