What this category's data looks like
Materials in the research ethics review scenario include ethics review applications, informed consent form templates, and risk assessment forms, with fields covering subject ID, research project name, review date, version number, and signature block. When these materials enter the model integration and configuration stage, the language model handles text dialogue and multilingual terminology extraction, the indexing model is responsible for chunking the terminology glossary and comparison drafts for retrieval, the reranking model optimizes terminology recall ranking, and speech synthesis and speech recognition handle oral records and interview audio respectively. The material source is the submitted drafts and receipts archived by the institutional ethics committee, and the update cadence is triggered by version revisions and annual continuing review. If the same term is written inconsistently across the application, consent form, and risk form, or if dates, quantities, or signature blocks are missing, it will directly enter the retrieval results.
Configuration and implementation checks
This step is executed by the server-side model channel configuration, outputting model binding relationships and channel responses. "Unified terminology glossary and comparison drafts" is then completed by subsequent AI nodes, code nodes, or manual work. Models are divided into five categories: language, indexing, reranking, speech synthesis, and speech recognition. Different stages bind different model types, and during configuration it is necessary to verify the type labels and the model names returned by the channel.
| Object | Check | Verification action |
|---|---|---|
| Language model | Corresponds to the dialogue stage; verify whether the type label is language model | In the model configuration, verify the type label and compare it with the model name returned by the channel to confirm that the dialogue and retrieval use this model |
| Indexing model | Corresponds to the retrieval stage; verify whether the type label is indexing model | In the model configuration, verify the type label, check the channel configuration and key source, and confirm that the type label is correct when retrieval reports an indexing model-related error |
| Informed consent form template | Fields must be consistent with the ethics review application and risk assessment form: subject ID, research project name, review date, version number, and signature block | Extract three terms, compare them with the same-named records in the application and risk form, and verify whether the name, quantity, date, and signature receipt are consistent |
| Risk assessment form | Fields must be consistent with the ethics review application and informed consent form template: risk level, review date, version number, and signature block | Extract three terms, compare them with the same-named records in the application and consent form, and verify whether the version number is synchronized and whether the signature or receipt is complete |
Three common errors
- Phenomenon: The same subject ID is written differently in the ethics review application and the informed consent form template. Cause: Multilingual terminology is not unified, translated names cannot refer back to the source text, and recall is misaligned during retrieval.
- Phenomenon: The review date and version number of the risk assessment form and the informed consent form template do not match. Cause: Versions are not synchronized, old drafts and new drafts are mixed into the same index, and retrieval results point to outdated materials.
- Phenomenon: The informed consent form template lacks a signature block or receipt record. Cause: Fields are not validated according to the category's verification dimensions, and missing items are not intercepted when materials enter retrieval.
How to verify the setup
Acceptance for this feature: run one dialogue and one retrieval, check the model identifiers in the responses, confirm that the binding relationships of the language model and indexing model are correct, and check whether the values of maxContext, maxTokens, and quoteMaxToken match the tokens of the quoted text; record the number of quotes separately.
Business test: sample three terms and check whether their spelling is consistent throughout the entire document
- In the model configuration, verify the type labels of the language model, indexing model, reranking model, speech synthesis, and speech recognition, and compare them with the model names returned by the channel.
- Extract three terms and compare them one by one against the name, quantity, date, version number, and signature receipt in the ethics review application, informed consent form template, and risk assessment form.