What the documents look like
Inbound receiving and putaway handles the first physical-and-record confirmation when goods enter the warehouse. Data comes from the WMS, ERP purchasing, supplier delivery notes and barcode scanning. The documents handled day to day are purchase arrival notes, inbound acceptance sheets, putaway tasks and discrepancy reports. Product and marketing copy drafting searches these materials for statements that can be cited. Retrieval results supply textual grounds only; they do not constitute business confirmation of a supplier, a storage location or an inventory state.
The smallest unit of search is the library. A collection is used only for classified management of data and has no bearing on search results. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries using different vector models cannot be associated together. A retrieval score reflects textual relatedness only. It does not carry permission or entity isolation, and it does not determine whether a record falls within the currently authorized business scope.
Inbound receiving materials update frequently, and drafting queries often mix precise elements such as model numbers, storage location codes and acceptance conclusions with natural-language phrasings. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, hybrid retrieval merges the two paths through RRF, and reranking reorders candidate text and returns a score from 0 to 1. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The reference cap controls, in tokens, the maximum amount of content cited per search. Question optimization performs coreference resolution and question expansion and makes an additional model call. Image retrieval requires image input and the corresponding model capability, and reranking requires a model enabled under the settings. The optional paths listed in the documentation do not mean that every request executes all of them.
A hypothetical paper exercise sample: in an e-commerce warehouse receiving scenario, the source text of the purchase arrival note reads "Purchase arrival note PO-2026-0417 states: the supplier delivered 200 units of Type A storage boxes; the receiving clerk completed the quantity count on 17 April 2026 and received 200 units." The inbound acceptance sheet YS-2026-0417-03 records "outer cartons intact, 10 units sampled with no damage, acceptance conclusion qualified." The putaway task SJ-2026-0417-11 assigns Type A storage boxes to location B-03-02 with status pending execution. The draft sentence to be assessed is: "This batch of Type A storage boxes has completed putaway and can be sold externally." The handover request to be checked is: "Please confirm, based on the above materials, the putaway completion state and sellable state of Type A storage boxes."
Constraints at the retrieval step
The retrieval step sits between these materials and the drafting task. It returns cited passages; it does not verify that a supplier delivered, that a location is free, or that stock is sellable. Several documented conditions shape how the step behaves.
Semantic retrieval converts the query and the stored content into vectors and compares distance. It handles paraphrases and cross-language phrasing but depends on model training and is affected by keyword and sentence completeness. Full-text retrieval matches keywords and suits exact queries such as numbers, model designations and proper nouns. Hybrid retrieval runs both and merges them through RRF, which usually broadens the candidate set. Because the merged range is large and cannot be filtered directly by similarity, a rerank model is typically used to reorder and to filter by its score. The minimum relevance value applies only under semantic retrieval or when reranking is used. It is a filter threshold, not a final ordering rule. When question optimization, hybrid retrieval, image retrieval or reranking is active, the final result may come from multi-path fusion and need not follow a single vector similarity order.
The reference cap is expressed in tokens rather than top k, because chunk lengths differ widely in mixed knowledge bases. Question optimization runs before retrieval and adds a model call, which can improve multi-turn retrieval while increasing latency. Image retrieval is not a separate capability outside the knowledge base search chain; it adds an image input path, and whether an image can be retrieved depends on whether a corresponding index was built at ingestion time. A plain text query with no image input does not exercise the image description or image vector paths.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only libraries containing inbound receiving documents that share one vector model | Check the vector model identifier of each library first, exclude those that differ, and do not run cross-model retrieval |
| Search mode | Run semantic, full-text and hybrid retrieval once each | Compare the cited content returned by the three paths for the same draft sentence, then decide which to use |
| Reference cap | Hypothetical starting value of 2000 tokens | Test with arrival notes and acceptance sheets of differing length, compare truncation against a smaller value |
| Minimum relevance | Hypothetical starting value of 0.3 | Label relevant and irrelevant samples first, then compare how many irrelevant passages are filtered at different thresholds |
| Question optimization | Enable for follow-up questions containing references, disable for complete draft sentences | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when hybrid retrieval returns many candidates | Compare the 0-1 score ordering before and after enabling, and check whether textual relatedness improves |
Three common errors
- Treating a retrieval score as a permission or entity filter. Input the draft sentence and the search parameters, compare whether the returned citations include unauthorized location or supplier records, and observe whether out-of-scope content appears. A similarity setting does not exclude unrelated businesses or unauthorized records.
- Writing a pending status as completed. Input the putaway task SJ-2026-0417-11 with status pending execution, compare it with the draft sentence "has completed putaway," and observe whether the two share the same subject, the same period and the same field with mutually exclusive values. A pending status does not establish that putaway is complete.
- Treating optional paths as executed on every request. Input a plain text draft sentence with no image, compare whether image description or image vector recall appears, and observe whether the retrieval chain follows only the text path. With no image input, image retrieval does not participate.
How to verify the setup
- Check the draft sentence against the hypothetical source text. Input the three source documents and the draft sentence, compare "has completed putaway" in the draft with "pending execution" in the putaway task, and observe whether they are mutually exclusive values of the same field. If the materials do not state a putaway completion time or completion confirmation, the statement is unsupported and the evidence gap is recorded.
- Check the sellable state against the handover request. Input the handover request and the three documents, compare whether the materials state a sellable determination or an authorized release field, and observe whether source text supporting that conclusion exists. When the materials do not state it, the matter remains pending verification and is not inferred as a denial or a conflict.
- Check downstream handling of an empty result. Input a paper empty citation array [], compare whether the downstream node continues along the path where output length is 0, and observe whether an interruption or exception appears. This check verifies only the empty-result handling design and does not claim that a fictional rule search returns [].