Data Characteristics
Deviation and CAPA (Corrective and Preventive Action) data in clinical trial pre-screening primarily includes unexpected event reports, root cause analysis documents, CAPA plans, implementation records, and verification reports. This data exists in a mixed format, combining structured data (e.g., database records) and unstructured data (e.g., PDFs, Word documents, scanned images). Data sources include Clinical Trial Management Systems (CTMS), Electronic Data Capture (EDC) systems, and Quality Management Systems (QMS). The update frequency depends on the clinical trial progress; deviation reports might be generated in real-time, while CAPA documents have a fixed lifecycle and approval process, leading to less frequent updates. Document structures typically include fields such as event description, occurrence time, responsible person, impact assessment, root cause, corrective actions, preventive actions, implementation progress, and effectiveness verification. Some fields may involve specific medical terminology, units of measurement (e.g., dose mg, duration days), and even scanned documents with handwritten annotations.
Deployment and Upgrade Constraints
The mixed structured and unstructured nature of Deviation and CAPA data requires FastGPT deployments to support multi-modal data processing, especially during document parsing and embedding generation. Large volumes of unstructured reports and scanned documents need efficient OCR and text extraction to ensure information completeness. Inconsistent update frequencies demand that the knowledge base supports incremental updates and version management to prevent data redundancy and ensure timely retrieval results. The presence of specific medical terminology and units of measurement means the model needs domain adaptability, potentially requiring fine-tuning or the use of domain-specific vocabularies to improve understanding accuracy. The deployment environment must consider compliance requirements for sensitive medical data, such as data encryption, access control, and audit logs, to ensure data security. Storing and retrieving large-scale unstructured documents places higher demands on underlying storage and computing resources, particularly in multi-node deployments, where data sharing and consistency become critical.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 500 MB | Deviation reports and CAPA documents may contain many images or scanned files, resulting in large file sizes. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Processing complex PDFs or scanned images for text extraction and OCR can be time-consuming. |
Chunk size | 800–1200 characters | Ensures each knowledge chunk contains sufficient context without being too long, which could lead to semantic drift. |
Recall count | Top 10 entries | Increases recall for pre-screening, covering more potentially relevant deviation or CAPA records. |
Similarity threshold | Calibrate by measurement | Balances recall and precision based on domain terminology similarity, avoiding false positives. |
Rerank result count | Top 5 entries | Selects the most relevant results after re-ranking, improving user experience. |
Common Pitfalls
- After knowledge base construction, query results are missing key information or fields are empty. This occurs when document parsing or OCR fails to accurately identify and extract all relevant content from unstructured documents, such as table data or handwritten annotations.
- In multi-node deployments, query results are inconsistent across different nodes. This happens due to improper knowledge base data synchronization or caching mechanisms, leading to different data versions across nodes.
- After adding a model, it is unusable or responds slowly. This occurs when the model version is incompatible with the FastGPT platform version, or the model lacks sufficient resources (e.g., GPU memory) during loading.
Verification Steps
- Upload typical deviation reports and CAPA documents. Verify that file parsing logs show no significant errors and that the knowledge base content preview accurately displays key information fields.
- After configuring and saving relevant parameters in the FastGPT administration interface, perform multiple queries via API or the interface. Check if the returned results are stable and meet expectations.
- In a multi-node environment, send the same query requests to different nodes. Compare the completeness and consistency of the returned results to ensure proper data sharing and synchronization.
Note: The values provided are common starting points. Measure them against your own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.