What the documents look like
Fruit-store procurement records centre on purchase orders, acceptance sheets, pesticide-residue test reports and loss records. The brief lists wholesale market prices, supplier quotations, store sales and test reports as data sources, with updates at daily or per-batch frequency. For copy drafting, these records supply the checkable wording a draft may use: grade, origin, test conclusion, loss basis. Compared with convenience-store fresh-food ordering, fruit procurement relies more on wholesale market conditions and grade acceptance, so retrieval terms frequently include exact fields such as grade names, batch numbers and test items. Semantic retrieval and full-text retrieval each have a place, and the choice depends on the question.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together; the model identifier is checked first and non-matching bases are excluded, without running a cross-model search to compare missed recall. Reference content is output as an array, and its length can be 0, so the empty-result path still completes. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not carry permission or entity separation. RRF is a fusion step for mixed results, not an on/off switch. Image retrieval requires image input and model capability; reranking requires a model and must be enabled by configuration.
Hypothetical drill sample: the source text to be checked reads, "Acceptance sheet, 12 June 2024: Shandong Red Fuji apples, grade one, 200 boxes arrived, 20 boxes sampled, pesticide-residue test report number JC-20240612-07, conclusion qualified." The proposed draft sentence reads, "This batch of Shandong Red Fuji apples has passed pesticide-residue testing, grade one." To be determined: whether "has passed pesticide-residue testing" and "grade one" in the draft sentence match the same batch and the same field of the acceptance sheet and test report above.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this run | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Purchase-order base, acceptance-sheet base, test-report base | All three use the same vector model; check the model identifier base by base before associating |
| Search mode | Hybrid retrieval | Grade and batch numbers need full-text hits, natural-language questions need semantic recall; compare the two result paths |
| Reference cap | 3000 tokens as the hypothetical starting value | Acceptance sheets and test reports have long passages; compare content completeness using materials of different lengths, then adjust |
| Minimum relevance | 0.4 as the hypothetical starting value | Label relevant and irrelevant samples first, then compare whether filtering drops key acceptance-sheet fields |
| Question optimization | Enabled | A multi-turn follow-up such as "the test conclusion for this batch" needs its subject completed; compare the original and optimized question to see whether the subject was altered |
| Result reranking | ReRank enabled | When candidates are numerous, rerank; compare how rerank scores and semantic scores order the same acceptance sheet |
Three common errors
- Treating minimum relevance as a permission or entity filter. The error appears as an assumption that raising the threshold excludes other stores or unauthorized records. The check is to enter the same question and compare the reference content returned before and after filtering, confirming that filtering changes relevance ordering only and not the accessible range.
- Associating knowledge bases that use different vector models. The error appears as placing bases with different vector models into one search node. The check is to verify each base's vector-model identifier, exclude non-matching ones, and record what is actually observed rather than inferring missed recall or redundant results.
- Reading "has passed pesticide-residue testing" in the draft sentence as all 200 boxes qualified. The error appears as a conclusion from 20 sampled boxes being written as a whole-batch conclusion. The check is to compare the subject scope of the draft sentence with the sampled quantity on the acceptance sheet, confirming whether the fields share the same subject, the same period and the same field.
How to verify the setup
- Enter the hypothetical source text and draft sentence above, and compare whether "has passed pesticide-residue testing" in the draft sentence matches the conclusion of test report JC-20240612-07 for the same batch. Expected observation: if the draft sentence does not limit the sampling scope, it is judged unsupported and needs "20 boxes sampled" added or a change to pending verification.
- Enter "what grade are these apples" and compare whether the returned reference contains the "grade one" field from the acceptance sheet. Expected observation: if the returned reference lacks that field or cites another batch, record the gap without claiming full recall.
- Enter a query unrelated to the acceptance sheet and compare the length of the returned reference array. Expected observation: if an empty array is returned, check how the downstream design handles empty results; do not claim that a fabricated rule search necessarily returns an empty array.