What this category's data looks like
In the academic journal business cooperation management scenario, materials usually come from three lines: business cooperation agreements, advertising review forms, and distribution cooperation materials, with update cadence following contract renewals, advertising schedules, and distribution receipts entering in batches. Field characteristics center on names, quantities, dates, and signature fields: the publication issues and print runs in agreements, the advertisers and placement periods in review forms, and the channels and receipt dates in distribution materials are often maintained by different handlers. When entering knowledge base retrieval and recall, the three types of materials each play different roles: business cooperation agreements provide the scope of cooperation and settlement basis, advertising review forms provide placement compliance and approval traces, and distribution cooperation materials provide channel quantities and receipt evidence. Together the three form the comparison draft for discrepancy localization, for subsequent nodes or humans to generate a discrepancy list accordingly.
Configuration and implementation checks
This step is executed by the knowledge base retrieval service (retrieval node and search test), outputting candidate snippets and similarity scores (after mode, merge, re-rank, filter, and limit truncation). The discrepancy list is then completed by subsequent nodes or humans.
| Object | Check | Verification action |
|---|---|---|
| Business cooperation agreement | Whether names, quantities, dates, and signature fields are consistent with the ledger | In the search test, search using key agreement fields, record the number of candidates and scores, and verify the signing date, print run, and signatures of both parties |
| Advertising review form | Whether the advertiser, placement period, and review signature are consistent with the agreement basis | Switch searchMode to compare semantic and full-text candidates, and see whether the placement start and end dates and the review signature fall in the same record |
| Distribution cooperation materials | Whether channel names, distribution quantities, and receipt dates are consistent with the agreement | Compare the ordering of the top few results before and after enabling usingReRank, and verify the receipt date and distribution quantity fields |
| Retrieval parameters | Whether candidates cover the key snippets of both materials | Use similarity to observe the score distribution, use limit to control the number of returned results, and compare names and dates item by item |
Three common errors
- Symptom: the name of the same record does not match between two materials. Cause: the agreement uses an abbreviation while the review form uses the full name, retrieval does not perform coreference resolution, and the candidate snippets are independent of each other.
- Symptom: quantities or dates do not match between two materials. Cause: versions are not synchronized, the old distribution materials remain in the knowledge base, and they are hit together with the new agreement during recall.
- Symptom: fields missing signatures or receipts are ignored. Cause: candidate snippets only capture the body text, the signature field and receipt date do not enter the cited content, and the discrepancy is not localized.
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 top few results; at the same time record the number of candidates and scores, and when the citation limit is involved, verify the tokens of the cited text, recording array item counts separately.
Business test: use one real record to find fields inconsistent with the ledger, hand them to a reviewer for confirmation, and write back.
- In the search test, search for the same record using the agreement name, the review form name, and the distribution material name respectively, and verify item by item the names, quantities, dates, and signature fields in the candidate snippets.
- Record the ordering of the top few results once before and once after enabling re-ranking, and use
similarityandlimittogether to judge whether the candidates cover the key fields of both materials.