What the documents look like
Opening support transfers operating capability. It is not a legal filing disclosure. The materials in this category are training attendance sheets, opening checklists, SOP manuals and assessment scores. They update per store, per month, so retrieval faces several versions of the same store across different months.
Each material type has a different object. A training attendance sheet records who was present. An opening checklist records materials and actions. An SOP manual records standard actions. An assessment score records how well a person performed. When a retrieval query contains only a store name, recalled content mixes material types together. Operating report review and exception review require attendance, checklist and score records for one store and one month to be compared under the same time scope. The retrieval step therefore needs to distinguish material type and period.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be linked together. During checking, confirm the vector model identifier first, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recall.
Search mode can be semantic search, full-text search or hybrid search. Compare each mode against the test question rather than assuming one mode suits all materials. The reference limit is measured in tokens. Use actual materials of different lengths to test whether content stays complete. Minimum relevance takes effect only in semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion. Compare the original question with the optimized question and check whether the subject was changed incorrectly. Hybrid search merges multiple result paths through RRF. RRF is a merging process, not a switch that can be turned on. Reranking requires a model and must be enabled in settings. Image retrieval requires image input and corresponding model capability. These optional paths do not mean every request executes all of them. Retrieval and relevance scores do not carry permission or entity isolation. A retrieval result is not an authorization.
Hypothetical paper test sample: the training attendance sheet for a franchise store in June 2025 reads "On 12 June 2025, 3 new cashier staff attended, 2 new inventory staff attended, supervisor Li signed to confirm." The opening checklist for the same month reads "On 20 June 2025, cashier counter materials complete, inventory counting process rectification not completed." The proposed verification sentence is "For this store in June 2025, both cashier and inventory staff training attended, and the opening check passed in full." The attendance sheet and the checklist share the same store and the same June 2025 period, but the checklist states that inventory counting process rectification was not completed. The phrase "passed in full" in the proposed sentence does not match the value recorded in the same field of the source text. The proposed sentence is unsupported.
Configuration and implementation checks
| Configuration item | Value or check for this page | Note |
|---|---|---|
| Linked knowledge bases | Select training, checklist, SOP and score bases that use the same vector model | Cross-model linking is not allowed; verify the identifier first |
| Search mode | Compare semantic, full-text and hybrid separately against the test question | Do not assume one mode suits all materials |
| Reference limit | For this hypothetical trial, start at a token amount covering the full text of one checklist | Adjust after comparing against actual materials of different lengths |
| Minimum relevance | Takes effect only in semantic search or result reranking | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare original and optimized questions; check the subject |
| Result reranking | Requires a model and must be enabled in settings | Rerank score is 0-1 and can be used for filtering |
Three common errors
- Treating a retrieval hit as authorization. Input "store X June 2025 checklist" and retrieval returns checklist content for that store. This only shows the content was recalled. It does not show the store passed the check or received opening approval. During checking, compare the retrieval result with the same field value in the checklist source text and observe whether wording such as "rectification not completed" conflicts with the proposed verification sentence.
- Linking knowledge bases with different vector models. Input two knowledge base identifiers with different vector model identifiers, compare whether the identifiers match, and exclude those that do not. Record the actual observation. Do not infer recall behavior after cross-model linking.
- Changing the subject through question optimization. Input "what is the second point." Compare the original question with the optimized question. If the optimized subject changes from inventory staff to cashier staff, the optimized result does not match the source subject. Record this and check it.
How to verify the setup
- Input the attendance sheet source text and the proposed verification sentence from the hypothetical sample. Compare "both cashier and inventory staff training attended" against the attendance numbers recorded in the attendance sheet and observe whether they match. Then compare "opening check passed in full" against "inventory counting process rectification not completed" recorded in the checklist and observe whether the same field values are mutually exclusive. If they are mutually exclusive, the proposed verification sentence is unsupported.
- Input two knowledge base identifiers with different vector model identifiers. Compare whether the identifiers match and observe whether they are excluded. Do not run cross-model retrieval and do not infer missed recall or redundant results.
- Input a clearly paper-based empty result []. Check how the downstream design handles empty references and observe whether the output path runs through. Do not claim that a fictional rule retrieval will always return [].