What this category's data looks like
This category's data involves three types of materials: budget adjustment applications, reimbursement forms, and contract invoices. The material fields include project name, budget subject, amount, quantity, date, and signature field. Sources are financial system exports, paper scans, and contract attachments, and the update cadence advances with approvals, changes, returns, and supplementary documents. When entering the knowledge base retrieval and recall function, budget adjustment applications play the role of comparing budget subjects and amounts before and after adjustment, reimbursement forms play the role of checking actual expenditures against the number of receipts, and contract invoices play the role of tracing suppliers, contract numbers, and invoice dates. Together, the three types of materials form the foundational corpus for the version ledger and change records, based on which the retrieval node recalls candidate fragments and similarity scores for subsequent nodes or manual organization of the current version and reference synchronization relationships.
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 version ledger and change records are then completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Budget adjustment application | Whether the budget subject, amounts before and after adjustment, approval date, and signature field are consistent with the contract invoice | Switch searchMode in the search test, and check whether candidates under semantic, full-text, or hybrid recall cover the amounts before and after adjustment and the approval date |
| Reimbursement form | Whether the reimbursement amount, quantity, occurrence date, and signature or receipt match the budget adjustment application | Test once each with synonym-rewritten questions, compare the number of candidates and scores, and record the original text fragments retained after similarity filtering |
| Contract invoice | Whether the supplier name, contract number, invoice date, and amount are consistent with the same record in the reimbursement form | Compare the ranking of the top few results before and after enabling usingReRank, and check whether the contract number and invoice date stably appear in the top candidates |
| Version ledger and change records | The current version is unique, old versions are traceable, references are synchronized, and key fields are still retained after reference count trimming by limit | After changing one piece of content, rerun retrieval to see whether the version number and references are updated synchronously; record the tokens of the reference text separately, and record the array count separately |
Three common errors
- Symptom: The name, quantity, and date of the same expenditure do not match between the reimbursement form and the contract invoice. Cause: The materials come from exports of different systems, the field definitions are not aligned, and retrieval recall merges fragments of different records by semantic similarity.
- Symptom: The budget adjustment application has been updated, but the knowledge base still references the old version's amount and approval date. Cause: The version is not synchronized; old versions are traceable but the current version is not unique, and the referenced content is not refreshed with the change records.
- Symptom: The reimbursement form lacks a signature or receipt, but retrieval still returns that fragment as a candidate. Cause: The signature field and receipt field were not validated when the material was ingested, and similarity filtering only trims by text relevance without judging signature status.
How to verify the setup
This function's acceptance: test once each with synonym-rewritten questions, compare the number of candidates and similarity scores, then toggle usingReRank to see whether the ranking of the top few results changes; switch searchMode to observe whether semantic and full-text candidates are merged; check the reference count after similarity filtering and the trimming result of limit, and record the tokens of the reference text separately.
Business test: change one piece of content, see whether the version number and references are updated synchronously, and confirm that the current version is unique, old versions are traceable, and references are synchronized.
- In the search test, run a synonym question once each for the budget adjustment application, reimbursement form, and contract invoice, record the number of candidates and scores, and mark fragments where the name, quantity, and date do not match.
- After changing one amount or date, rerun retrieval, check whether the version number increments, whether the old version is traceable, and whether the referenced content is updated synchronously, and check whether the signature and receipt fields are complete.