What the documents look like
Store count and loss control work centres on book-to-physical differences and shrinkage. The working materials are count sheets, variance analyses, loss-report approvals and surveillance review records. These sit apart from supplier settlement, which handles accounts payable rather than stock discrepancies.
Materials are refreshed on a monthly or quarterly cycle. In cross-border operations the same count package can exist in a Chinese version and a foreign-language version at the same time, so retrieval has to locate the variance note for one SKU across languages. Variance investigation depends on exact identifiers such as SKU numbers, location codes and document numbers. Two phrases that read as similar in meaning do not necessarily point to the same record. This is why the search mode has to be chosen according to the type of question, and why a low similarity score cannot be treated as a substitute for checking which record was actually retrieved.
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, so the vector model identifier of each base is checked first and any base that does not match is excluded. Search modes are semantic search, full-text search and hybrid search. Hybrid search merges multiple result paths through RRF. Reranking requires a model and runs according to its configuration. The citation limit is expressed in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimisation performs coreference removal and question expansion and makes an additional model call. Retrieval itself does not decide authorisation. Whether retrieved content may be used for a loss-report approval or an external handover depends on the authorisation state of the corresponding business step, and a pending review is not an approval.
Hypothetical drill sample: a store loss-prevention team holds a paper variance analysis whose text reads, "March 2025 count, location A-12, SKU 8842, book quantity 120 units, physical count 96 units, variance 24 units, cause under investigation, loss-report application not yet submitted for approval." A separate draft sentence is to be assessed: "This SKU has completed loss-report approval and has been written off at 24 units." The check compares the approval state in the source text with the completion state claimed in the draft sentence. The two take mutually exclusive values on the same object, the same period and the same field, so the draft sentence is not supported by the source text.
Constraints at the retrieval step
The association rule constrains the first decision. A base whose vector model identifier differs from the others is excluded before any search runs. Cross-model retrieval is not performed in order to compare what might be missed.
Search mode selection follows question type. Semantic search computes distance between vectors and supports cross-language matching, so a Chinese question can match an English knowledge point. Full-text search depends on keyword hits and suits identifiers, model numbers, proper nouns and error codes. Hybrid search uses both and merges them through RRF. Because the merged candidate pool is wide and cannot be filtered directly by similarity, reranking is normally used afterwards and its score is used for filtering. Reranking requires a model and runs according to its configuration.
The citation limit caps how many tokens of content are cited per search. It is expressed in tokens rather than a top-k count because chunk lengths vary widely in mixed knowledge bases. The minimum relevance value is a threshold between 0 and 1 that removes low-relevance results, and it applies only in semantic search or when reranking is used. It is a filter threshold, not a final ordering rule. When question optimisation, hybrid search, image retrieval or reranking is active, the final result may come from a merged multi-path recall and does not necessarily follow a single vector similarity order.
Question optimisation runs before retrieval and adds a model call. It improves retrieval in multi-turn dialogue and expands semantic richness, and it also adds latency. Image retrieval depends on whether a vision model is configured or whether the vector model accepts image input, and on whether an image index was built at ingestion time. Neither path runs on every request.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One Chinese count-material base | Check the vector model identifier of each base first, associate only bases with matching identifiers, and do not run cross-model retrieval |
| Search mode | Full-text search for identifier questions, semantic search for natural-language questions | Test the same set of SKU numbers and descriptive phrasings separately and compare hit results |
| Citation limit | 2000 tokens | A starting value for this drill; compare whether citation content is truncated using count-sheet fragments of different lengths |
| Minimum relevance | 0.5 | A starting value for this drill; label relevant and irrelevant samples first, then compare filtered results |
| Question optimisation | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimised question and check whether subjects such as SKU and location were altered |
| Result reranking | Enabled when candidates are numerous | Compare ordering before and after enabling and check whether the score is used for filtering |
Three common errors
- Treating the minimum relevance value as an authorisation or entity filter. It is a threshold between 0 and 1 and does not distinguish stores, departments or authorisation scope. Retrieved content still requires a business-side check of its authorisation state.
- Describing RRF as an independent switch that can be turned on. RRF is the merge process for multiple result paths inside hybrid search and is not a configuration item. Reranking is the part that requires a model and runs according to its configuration.
- Treating a retrieval hit as a completed approval. Retrieving a draft loss-report application does not mean the application was submitted or approved. A pending-review state and an approved state are different facts.
How to verify the setup
- Enter the hypothetical source text and the draft sentence given above. Compare the approval-state field in the source text with the completion state claimed in the draft sentence. If the two take mutually exclusive values on the same object, the same period and the same field, the draft sentence is judged unsupported and is recorded as pending verification.
- Enter SKU 8842 and location A-12, and run full-text search and semantic search separately. Compare whether the returned citation content contains that variance record. If one mode returns nothing, record that observation and check whether the data has completed training. Do not assume the first result must be a hit.
- Enter a question that clearly has no corresponding material and compare the length of the returned citation array. If the length is 0, check whether the downstream step continues with an empty citation, and do not assert that a fabricated rule necessarily returns an empty array.