What this category's data looks like
This page addresses the scenario of signing research project task books; the materials include the project task book, the subject task breakdown table, and the budget book. The task book provides the overall objectives and the signature page, the breakdown table carries the subject names, quantities, start and end dates, and responsible persons, and the budget book lists the account amounts and dates. When these three types of materials enter knowledge base retrieval and recall, each participates as an independent source in candidate fragment generation: the task book and breakdown table lean toward semantic paragraphs, while the budget book leans toward table fields. The update cadence is version-controlled; each signing or revision produces a new version, and old versions are retained for comparison. Field characteristics center on names, quantities, dates, signatures, and receipts, and the same record must remain consistent across multiple materials.
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, re-rank, filter, and limit trimming). The handover-ready deployment status and operation records are completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Project task book | The name, date, and signature page fields must be aligned item by item with the breakdown table and budget book, and the version number and receipt status must be consistent | In the search test, retrieve the task book title and signature page, record the number of candidates and scores, and check searchMode and similarity |
| Subject task breakdown table | The subject names, quantities, start and end dates, and responsible persons must match the task book and budget book; when versions are not synchronized, the latest version prevails | Switch searchMode to hybrid, compare whether the candidates contain both semantic and full-text results, and record the number of candidates after limit trimming |
| Budget book | The account amounts, dates, and signature receipts must be consistent with the task book and breakdown table, and the name and quantity of the same record must not differ in two places | Turn usingReRank on and off and retrieve once each, compare the order of the first few items, and record the similarity scores and cited text tokens |
Three common errors
- Symptom: The same subject name is written differently in the task book and the breakdown table, and retrieval candidates do not match. Cause: The two entries were not field-aligned, and the versions were not synchronized.
- Symptom: The budget book date is inconsistent with the task book signing date, and the handover form lacks a record of this item. Cause: The old version was not replaced after revision, and the receipt was not archived.
- Symptom: There are many candidates, but the ones ranked first are all irrelevant, and the answer lacks key paragraphs. Cause: Re-ranking was not enabled or the similarity filter is too tight, and the citation limit trimming is excessive.
How to verify the setup
Acceptance of this feature: test once each with synonym-rewritten questions, compare candidates and scores, then toggle re-ranking on and off to see the ordering of the first few items.
Business test: check item by item against the handover form once, and see which record is missing.
- In the knowledge base search test, switch
searchMode, record the number of candidates and scores, and checksimilarityandlimit. - Compare the order of the first few items before and after enabling
usingReRank, and separately record the number of items in the citation array and the cited text tokens.