Data Characteristics
Pharmacovigilance data primarily originates from post-market surveillance reports, adverse drug reaction (ADR) reports, drug inserts, clinical trial reports, academic papers, and regulatory guidelines. This data updates frequently. Post-market surveillance and ADR data, in particular, may update quarterly or even monthly. Document structures vary, including unstructured free-text reports, semi-structured tabular data (e.g., CIOMS I forms), and structured database records. Fields cover patient demographics, drug information (batch number, dosage, usage), adverse event descriptions (ICD-10 codes), event onset and outcome, and causality assessment. Units typically include dosage units (mg, g, IU), frequency units (times/day, times/week), and time units (hours, days, months). These units may appear in non-standardized forms across different reports.
Constraints on Knowledge Base Retrieval and Recall
The high update frequency of pharmacovigilance data requires the knowledge base to support efficient incremental updates and version management. This ensures the timeliness of retrieval results. Diverse document structures mean the knowledge base must parse and embed various file formats. Extracting key information from free text is crucial. Semi-structured tabular data has strong field interdependencies; the RAG system needs to understand semantic relationships between fields to avoid limitations of single keyword matching. Non-standardized unit expressions challenge entity recognition and normalization, potentially leading to missed information due to unit mismatches during retrieval. Additionally, ADR reports often contain numerous medical terms and abbreviations. The knowledge base embedding model must possess domain-specific knowledge to improve retrieval accuracy and recall.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 800–1200 characters | Pharmacovigilance documents are often lengthy reports. Increasing chunk size helps retain context and reduces information fragmentation. |
Chunk Overlap Length (Overlap Length) | 100–200 characters | Ensures contextual continuity at chunk boundaries, improving recall for information spanning multiple chunks. |
Recall count (Recall Count) | 8–12 items | Adverse event reports may involve multiple related factors. Increasing the recall count can cover more potential associated information. |
Similarity threshold (Similarity Threshold) | Calibrate based on actual measurements (suggested 0.7–0.8) | Adjust based on the specific embedding model and data characteristics. Aims to balance recall and precision. |
Rerank result count (Reranked Return Count) | Top 5 items | After initial recall, use a reranking model to further optimize results and select the most relevant few items. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Pharmacovigilance report files can be large, and parsing may take a long time. A longer timeout is needed. |
Common Pitfalls
- When importing Markdown files into the knowledge base, an
Invalid array lengtherror often indicates excessively large file content or an abnormal format. This prevents the parser from correctly segmenting the file. - A configured knowledge base may fail to recall relevant information during actual Q&A. This could be due to the embedding model lacking domain-specific knowledge, preventing accurate understanding of pharmacovigilance terminology.
- Knowledge base search may work in debug preview but fail after integration into an external application. This could be due to incorrect API key configuration or restricted network access permissions.
Verification Steps
- Upload various types of pharmacovigilance documents (e.g., ADR reports, drug inserts). Observe file parsing status to ensure all files process successfully.
- For specific adverse events or drugs, construct queries containing specialized terminology. Check if retrieval results accurately recall key information and assess the completeness of recalled documents.
- Use queries containing non-standardized units or abbreviations. Verify if the knowledge base correctly identifies and matches relevant documents. For example, query "twice daily" and "bid" to find corresponding reports.
The values provided are common starting points and should be measured against the reader's 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.