Knowledge Base Retrieval for University Laboratory Safety Classification: Trace Exceptions and Review Results

What this category's data looks like

Materials for safety grading and classification of university laboratories usually include three types: laboratory grading and classification tables, hazard source lists, and safety responsibility statements. The grading and classification table gives levels and categories by room and laboratory, with fields including name, quantity, date, and signature; the hazard source list registers hazard source name, quantity, storage location, and responsible person item by item; the safety responsibility statement carries the responsible person and receipt signature. The material source is submission by each department and laboratory, and the update rhythm rolls with semesters, project changes, and rectification receipts. When entering this function, the three types of materials are each split and stored as knowledge base documents; the grading and classification table serves as the basis for level determination, the hazard source list serves for quantity and location verification, and the safety responsibility statement serves for signature and receipt confirmation. The retrieval service is only responsible for recalling candidate fragments and similarity scores, while the anomaly list and review conclusions are completed by subsequent nodes or manually.

Configuration and implementation checks

This step is executed by the knowledge base retrieval service (retrieval node and search test), outputting candidate fragments and similarity scores (after mode, merge, rerank, filter, and upper-limit trimming). The "anomaly list and review conclusions" are completed by subsequent nodes or manually.

Three common errors

• Phenomenon: The name and quantity of the same record do not match between the grading and classification table and the hazard source list, and during review it is impossible to determine which one prevails. Cause: The two materials are entered by different submitters, their update rhythms are not synchronized, and unified field validation is lacking.
• Phenomenon: The date on the safety responsibility statement is inconsistent with the date on the grading and classification table, and a missing receipt is still treated as confirmed. Cause: Versions are not synchronized, the old version of the responsibility statement was not updated with the rectification receipt, and the signature column was left blank.
• Phenomenon: There are many candidates, but the ones ranked first are all irrelevant, and the answer lacks key paragraphs. Cause: The retrieval mode does not match the material form, rerank is not enabled, and key fragments are squeezed out after similarity filtering and citation upper-limit trimming.

How to verify the setup

This function's acceptance: test once each with synonym-rewritten queries, compare candidates and scores, then toggle rerank to inspect the ranking of the first few items; record the number of candidates and scores, switch searchMode to observe differences between semantic and full-text candidates, and verify whether the fragments retained after similarity filtering contain key fields.

Business test: create an anomaly record and see whether it can be traced back to the source material; verify the fields, dates, and signatures of the grading and classification table, hazard source list, and safety responsibility statement, confirm that the anomaly is supported by original records and has the reviewer's signature, and then hand it to the responsible person for handling.

• In the knowledge base search test, retrieve once each for the same question using synonym rewrites, record the number of candidates and scores, and switch searchMode to compare semantic, full-text, and hybrid results.
• Retrieve once each with usingReRank enabled and disabled, compare the ranking of the first few items; after filtering by similarity, verify the tokens and array length of the cited text, and record them separately.