Data Characteristics in This Domain
Site Management Organizations (SMO) in pharmacovigilance primarily generate data internally from clinical trial sites. Data types include clinical trial protocols, subject informed consent forms, subject medical records, adverse event (AE) reports, serious adverse event (SAE) reports, laboratory test results, and safety update reports. These documents are typically in PDF, Word, or structured data formats (e.g., CSV, XML). Adverse event reports are updated frequently, possibly daily or weekly, while protocols and informed consent forms are updated less often. AE reports often contain fields such as event description, occurrence time, severity, outcome, related drugs, and medical history. Some fields may use medical terminology or coding, such as MedDRA codes. Laboratory test results include numerical values, units, and reference ranges.
Constraints on Citation and Traceability
The diversity of SMO data sources requires the system to handle multiple file formats. The frequent updates of adverse event reports demand real-time knowledge base capabilities, with an update mechanism that supports incremental synchronization to avoid full rebuilds. Medical terminology and coding within documents require the language model to accurately identify and understand their semantics, which impacts text embedding quality. The presence of numerical values and units necessitates accurate handling of numerical ranges and unit conversions during citation. The structured nature of adverse event reports makes it possible to cite and trace specific fields. The system must be able to parse this structured information and precisely point to the field's location in the original text during citation.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 500 MB | Clinical trial documents, especially PDFs with images or extensive charts, can be large. |
Chunk size (Segment Length) | 800 characters (characters) | Balances the descriptive nature of adverse event reports with the paragraph integrity of other documents, preventing truncation of critical information. |
Chunk Overlap Length (Segment Overlap Length) | 100 characters (characters) | Ensures contextual continuity, especially when processing medical terminology or long sentences. |
Recall count (Recall Count) | Top 10 entries (top 10) | Considering the detail and diversity of adverse event reports, increasing the recall count covers more potential related information. |
Similarity threshold (Similarity Threshold) | 0.75 | Balances recall and precision, avoiding retrieval of irrelevant documents while ensuring critical information is not missed. |
Rerank result count (Rerank Return Count) | Top 5 entries (top 5) | After reranking, filters for the most relevant and representative citation entries, reducing redundancy. |
Common Mistakes
- The documents cited in the response do not match the question. Investigation revealed that the knowledge base update mechanism did not synchronize the latest adverse event reports in time, leading to citations of outdated information.
- The model could not accurately answer questions about the severity of specific adverse events. Examination showed that the text content extraction component failed to correctly parse the
severityfield from structured adverse event reports. - The chat interface did not display the page or specific paragraph of the cited document. Debugging revealed that the
show_source_documentparameter was not enabled in the production environment, hiding citation information.
How to Verify Configuration
- Verify that the update date of adverse event reports in the knowledge base aligns with the update frequency of the original data source, ensuring information timeliness.
- Through simulated questioning, check if the system accurately understands specific medical terminology and codes in adverse event reports, and confirm that citations point to the corresponding fields in the original text.
- Check if the document links in the response are clickable and accurately navigate to the corresponding page or paragraph in the original text, confirming that the traceability function works correctly.
- For laboratory test results containing numerical values and units, verify that the values and units are correct when cited and reflect their contextual relationship.
Note: The values provided are common starting points and should be measured 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.