Reranking and Hybrid Retrieval for Graduate Education Process Management: Find Supporting Evidence

Graduate Education: review reranking for evidence lookup. Check source versions, input mappings and the values received by downstream nodes.

What this category's data looks like

Graduate training process management involves three core types of materials: training plans, proposal reports, and mid-term assessment forms, which are submitted by supervisors, colleges, and the graduate school at the enrollment, proposal, and mid-term milestones respectively. A training plan contains course names, credits, start and end dates, and the supervisor's signature; a proposal report contains the topic title, proposal date, review comments, and signature; a mid-term assessment form contains the assessment conclusion, material checklist, date, and signature. Materials are updated once per semester or per academic year, and the version archived by the college is the current one. When these three types of materials enter hybrid retrieval, the training plan and proposal report provide the basis for courses and topics, while the mid-term assessment form provides the basis for process assessment, jointly supporting cross-material business basis queries.

Configuration and implementation checks

The rerank model in the retrieval service only outputs the sorted candidates and rerank scores; the basis entries that can be traced back to the original text must then be handled by citation trimming and answer generation. Hybrid retrieval simultaneously runs the vector and full-text paths and merges the results; after merging, the scope is large and similarity filtering cannot be applied directly, so the rerank model is usually then used to sort and filter by rerank score.

ObjectCheckVerification action
Training planCourse names, credits, start and end dates, and supervisor signature must be consistent with the corresponding fields in the proposal report and mid-term assessment formCheck the two-path switches of searchMode, the similarity threshold, and the usingReRank binding; compare the number of candidates and scores before and after reranking
Proposal reportTopic title, proposal date, review comments, and signature must correspond to the records in the mid-term assessment formCheck the limit number of candidates and usingReRank score filtering; see whether the top few come from the merge of the two paths
Mid-term assessment formAssessment conclusion, material checklist, date, and signature must be synchronized with the versions of the training plan and proposal reportCheck the number of candidates after similarity filtering and the rerank scores; confirm that the number of citation-trimmed entries and the citation text tokens are recorded separately
Rerank modelAlready bound and used for sorting and filtering, outputting the sorted candidates and rerank scoresCompare the order and scores of the top three before and after reranking; confirm that the candidates come from both the vector and full-text paths

Three common errors

  • Symptom: The name, credits, or dates of the same course do not match across two materials. Cause: The versions of the training plan and proposal report are not synchronized, and archived copies are mixed with submitted copies.
  • Symptom: The mid-term assessment form lacks the supervisor's signature or the college receipt, and entries cannot be traced back to the original text. Cause: Signatures were not completed during the material circulation process, and the version was not synchronized to the current one.
  • Symptom: There are many candidates but the top few are irrelevant, and similarity filtering appears to have failed. Cause: The scope after merging is large, and the rerank model was not enabled or filtering by rerank score was not applied.

How to verify the setup

Acceptance of this feature: compare the order and scores of the top three before and after enabling reranking, check the status of usingReRank, similarity, limit, and searchMode, and record the number of candidates and scores.

Business test: ask a cross-material basis question, see whether the answer gives traceable sources, and check that the number of citations and the citation text tokens are recorded separately.

  • In the search parameters, check the vector and full-text path switches, confirm that both paths of searchMode are enabled, and see whether candidates come from both paths.
  • Compare the ranking and scores before and after reranking, check the usingReRank binding and similarity filtering, and confirm that the top few are relevant to the business basis.