What the documents look like
Opening support is a transfer of operating capability, so its materials center on training attendance sheets, opening checklists, SOP manuals, and assessment scores. Sources include the training system, the supervisor app, store photographs, and the headquarters SOP library. Records update per store and per month. One SOP manual commonly exists as a headquarters original alongside multilingual translations, while attendance sheets and checklists are produced separately for each store and each batch.
When these materials enter retrieval, the subject, object, state, and validity scope are often spread across different passages. An attendance sheet records participants and dates, a checklist records materials and store status, an SOP manual records required actions, and an assessment score records an individual result. Retrieval must associate these fields with a specific store and batch before later checking can rely on them.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model. Libraries using different vector models cannot be associated together. The practical step is to check the identifiers, exclude those that do not match, and record what is actually observed. No conclusion about missed recall or redundancy follows from that check.
Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval relies on keyword hits and suits numbers, models, and proper nouns. Hybrid retrieval uses both and merges the results through RRF. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds one model call. Reranking requires a model and must be enabled by configuration. Image retrieval requires an image input and the corresponding model capability. These paths do not run on every request.
Semantic recall and relevance scores do not carry permission or entity isolation. The same-vector-model requirement for associated libraries does not extend to external systems. Retrieval itself is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization each need their own downstream implementation.
Hypothetical paper exercise sample: a franchisee training file contains an "Opening Support Handover Note" whose original text reads: "The opening support materials for this batch of stores were compiled by the regional training department in March 2024. Attendance sheets and checklists are filed by store number. The SOP manual includes a Chinese original and an English translation. Assessment scores were exported from the training system and have not yet completed review." A draft sentence awaits judgment: "The materials for this batch have completed review and can be used directly for the opening support handover."
Configuration and implementation checks
| Configuration item | Relevant setting for this page | How to check |
|---|---|---|
| Associated knowledge bases | Select training archive libraries that use the same vector model | Compare vector model identifiers across libraries and exclude mismatches |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | Run the same question in all three modes and compare recalled content |
| Citation limit | For this hypothetical trial, start at the token amount needed to cover one complete SOP manual passage | Compare against materials of different lengths for content completeness |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Observe whether relevant samples survive filtering |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Requires a model and must be enabled by configuration | Inspect rerank scores and candidate order |
The hypothetical starting value for the citation limit is chosen because SOP manual passages vary considerably in length. It should be compared against shorter checklist entries and longer manual passages before adjustment.
Three common errors
- Associating libraries that use different vector models. Input: two libraries with different vector model identifiers. Compare whether the identifiers match. The observable result is a mismatch, and the library should be excluded. No claim about missed recall or redundancy follows.
- Treating minimum relevance as a permission or entity filter. Input: a checklist entry from an unrelated store. Compare whether it is still recalled after filtering. The observable result is a low relevance score without exclusion, which shows the value does not perform isolation.
- Treating "has completed review" in the draft sentence as an accomplished fact. Input: the sample original text and the draft sentence. Compare "have not yet completed review" with "have completed review." The observable result is that the same field holds mutually exclusive values, and the draft sentence is unsupported.
How to verify the setup
- Input the sample original text and the draft sentence. Compare "have not yet completed review" in the original with "have completed review" in the draft. Under the same subject, same period, and same field, the values are mutually exclusive, so the draft sentence is unsupported and the evidence gap must remain.
- Input two libraries with different vector model identifiers. Compare whether the identifiers match. When they do not match, exclude the library and do not run cross-model retrieval to compare missed recall.
- Input a result that is explicitly empty, meaning a paper input of []. Check whether the downstream design accepts a citation output of length zero. Do not claim that a fictional rule-based retrieval necessarily returns [].
All three verification steps address the primary task on this page. Expected results depend on actual observation, and no test has been claimed as performed.