What the documents look like
Fresh-food supplier claims pursue responsibility for quality at arrival. They do not handle natural in-store shrinkage. The main data sources are the quality inspection system, WMS receiving records, supplier contracts, and financial reconciliation. The documents handled day to day are quality exception sheets, photo evidence, deduction notices, and claim reconciliation statements. These materials update per batch and per month. One arrival batch may carry three documents at once: a receiving record, an exception sheet, and a deduction notice. The three may not agree on the same field.
Copy drafting for goods and marketing draws citable wording from these documents. Whether a draft sentence holds depends on whether the retrieved original text states the corresponding subject, object, status, and valid scope. Retrieved content is candidate material. Whether a sentence is supported by the documents still requires manual comparison.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. Check the vector model identifier first, exclude bases that do not match, and do not run cross-model retrieval to compare missed recalls. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Numbered items, models, and proper nouns suit full-text or hybrid retrieval. Natural-language questions suit semantic or hybrid retrieval. The reference cap is measured in tokens. Minimum relevance takes effect only in semantic retrieval or when result reranking is used. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process, not a switch to turn on. Reranking requires a ReRank model and must be enabled in settings. Empty reference output may have a length of 0, and the downstream path still runs.
Hypothetical paper exercise: a fresh-food supplier contract annex states "Arrival batch 2024-05-10; spinach pesticide residue test result exceeds the limit agreed in the contract; deduction notice number KC-20240510-03 has been generated; claim reconciliation statement awaits supplier confirmation." The draft sentence under review is "This batch of spinach failed inspection, and the deduction has been completed under the contract." The comparison must set the draft sentence against the original text. The original states that the test exceeded the limit, that a deduction notice has been generated, and that the reconciliation statement awaits confirmation. The draft sentence says the deduction "has been completed," which conflicts with the original wording "awaits supplier confirmation." The sentence is not supported by the material.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Quality exception sheet base, deduction notice base | Check whether the two bases share the same vector model identifier; associate them only if the identifiers match |
| Search mode | Hybrid retrieval | The sample contains number KC-20240510-03, which full-text retrieval can match, and natural-language description, which semantic retrieval can match |
| Reference cap | Set to 2000 tokens as the hypothetical starting point for this exercise | Test with actual materials of different lengths to see whether content stays complete; compare against a shorter cap before adjusting |
| Minimum relevance | Set to 0.5 as the hypothetical starting point for this exercise | Label relevant and irrelevant samples first, then compare filtering results at different thresholds |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enabled in settings | A ReRank model must be available; compare candidate ordering before and after reranking |
Three common errors
- Treating retrieval results as authorization. Retrieval returns the original deduction notice text. That does not mean the copy may state "the deduction has been completed." The original states "awaits supplier confirmation," while the draft sentence says "has been completed," which is unsupported. Check: input the draft sentence and the retrieved original text, compare the values of the same field, and observe whether mutually exclusive values exist.
- Associating bases that use different vector models. Check: input the candidate knowledge base list, compare the vector model identifier of each base, exclude bases whose identifiers do not match, and record the actual observation.
- Treating minimum relevance as a permission or entity filter. Minimum relevance is a filtering threshold. It does not exclude unrelated business or unauthorized records. Check: input an unrelated query and observe whether the returned result is empty. An empty result cannot be inferred to mean permission isolation is active.
How to verify the setup
- Input the hypothetical sample original text and the draft sentence "This batch of spinach failed inspection, and the deduction has been completed under the contract." Compare "awaits supplier confirmation" in the original with "has been completed" in the draft sentence, and observe whether the unsupported status is identified. Condition met: the draft sentence is marked as pending verification and is not marked as confirmed.
- Input the candidate knowledge base list, compare the vector model identifier of each base, and observe whether bases with non-matching identifiers are excluded. Condition met: only bases with matching identifiers remain, and excluded items are recorded.
- Input a natural-language question that contains no number, observe the candidate content returned by hybrid retrieval, and compare the relevance of the candidate content to the question. Condition met: the candidate content includes a relevant passage. First-item hit and full recall are not guaranteed.
The following public technical facts were reviewed for identifying related content in the documents; the body is written independently and is not required to copy them as a fixed five-row table: [["Associated knowledge bases", "The knowledge base search node can associate one or more knowledge bases that use the same vector model."], ["Search mode", "Compare semantic, full-text, and hybrid retrieval against the test question; do not assume one mode suits all materials."], ["Reference cap", "The unit is tokens; test with actual materials of different lengths to see whether content stays complete, then adjust the cap."], ["Minimum relevance", "Takes effect only in semantic retrieval or when result reranking is used; label relevant and irrelevant samples before adjusting the threshold."], ["Question optimization", "Used for coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered incorrectly."]]