What the documents look like
Opening support is a transfer of operating capability, so its records sit closer to daily store execution than to legal filing. The materials named for this category are training sign-in sheets, opening checklists, SOP manuals, and assessment scores, refreshed per store and per month. Procurement and supplier information usually reaches retrieval as an attachment to an SOP manual, a checklist line, or a training handout, which means ordering steps, receiving verification, and supplier contact details can appear inside one store's document set.
The practical difficulty is that subject, store, period, and version are easy to mix. A checklist line about a shelf arriving and a sign-in sheet about cashier training belong to the same store and month, yet neither one states that supplier documents were handed over. Retrieval works on the store, month, and field actually written in the source; joining records across stores or across months produces a statement the source never made. Update frequency by itself does not change how retrieval behaves, and this page draws no conclusion about batch synchronization or real-time mechanisms from it.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together, so the check is to read the vector model identifier and exclude those that do not match, rather than running a cross-model search to compare what was missed. Search mode is chosen by testing the actual question against semantic retrieval, full-text retrieval, and hybrid retrieval; no single mode is assumed to suit every procurement document.
The citation limit is measured in tokens. Because chunk lengths differ greatly between a short checklist line and a long SOP section, the limit is tested with real documents of different lengths to see whether content is cut off. The minimum relevance value is a number between 0 and 1, and it takes effect only under semantic retrieval or when result reranking is used, so related and unrelated samples are labelled first and the threshold is compared against them. Question optimization performs coreference resolution and question expansion; the original question and the optimized question are compared to confirm that store and month subjects were not altered. A similarity score measures relevance and carries no permission or entity-isolation function, and retrieval results do not constitute authorization. The node outputs cited content as an array whose length may be zero, and the output path still runs when nothing is found.
Hypothetical paper exercise: a franchise store's March 2024 training sign-in sheet records "March 5, cashier position training, four attendees signed in." The same month's opening checklist records "March 6 inspection, shelving in the stock area in place, supplier contact sheet not attached." Version 2 of the SOP manual states "complete system entry within 24 hours of receiving goods." The draft sentence under review is "the store completed the supplier document handover in March." The source text covers sign-in, inspection, and an SOP clause; it contains no record that the supplier document handover was completed, so the draft sentence is unsupported and the evidence gap is recorded.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the store SOP base that uses the same vector model | Read the vector model identifier first, exclude bases with a different model, and do not run a cross-model search |
| Search mode | Run the sample question once under semantic retrieval and once under full-text retrieval | Compare which of the sign-in sheet, checklist, and SOP clause are hit in each run |
| Citation limit | Set a token amount that covers one complete checklist entry | Test checklist entries of different lengths to see whether content is truncated |
| Minimum relevance | Label related and unrelated samples first, then adjust the threshold | Compare filtering results only where semantic retrieval or reranking is in effect |
| Question optimization | Enable for follow-up questions containing references, disable for explicit questions | Compare the original and optimized question to confirm store and month subjects were not altered |
| Result reranking | Enable when candidates are numerous and compare rerank scores | Reranking requires a model and must be enabled by configuration; it is not treated as on by default |
Three common errors
- Treating a similarity score as a permission filter. The input is one supplier document query, the comparison is between score level and record ownership, and the observable result is that a low-scoring record can still belong to another store; the score does not change authorization state.
- Treating a pending state as approval. The input is the checklist line "supplier contact sheet not attached," the comparison is between a completed handover record and a pending item, and the observable result is that the completion field is absent, so the handover cannot be judged complete.
- Linking knowledge bases that use different vector models. The input is the vector model identifier of each base, the comparison is whether the identifiers match, and the observable result is that a mismatch leads to exclusion, with no cross-model search performed.
How to verify the setup
- Input the sign-in sheet, checklist, and SOP clause from the hypothetical exercise, then compare whether the retrieved results cover the same store and the same month. The condition is met when all three document types are recalled with a consistent subject.
- Input the draft sentence "the store completed the supplier document handover in March" and compare the retrieved source text for a handover completion field. The condition is met when the field is absent, so the sentence is judged unsupported.
- Input a paper empty result
[]and compare whether the downstream path still runs. The condition is met when cited content has length zero and the flow continues; no claim is made that a fabricated rule search always returns an empty array.