Data Characteristics
Pharmacovigilance data originates from adverse drug reaction (ADR) reports, clinical trial data, literature, regulatory documents, and various guidelines. This data updates frequently, especially ADR reports, which may update daily or weekly. Document structures vary, including structured report forms, semi-structured medical texts (e.g., case reports, pharmaceutical assessments), and unstructured regulatory documents and research papers. Documents often contain fields such as patient demographics, drug information (batch number, dosage form, dose), ADR event descriptions, diagnostic results, treatment measures, and outcomes. They involve medical terminology, units of measurement (e.g., mg, ml, μg/kg), timestamps (e.g., ISO 8601 format), and may include mixed languages.
Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts
High update frequency of pharmacovigilance documents requires the knowledge base to quickly synchronize the latest information, ensuring real-time relevance and accuracy of conversation content. Diverse document structures, particularly the semi-structured and unstructured nature of medical texts, demand higher quality knowledge chunking and vectorization. This requires more refined segmentation strategies to prevent semantic loss. Complex medical terminology and units of measurement, along with potential colloquial descriptions, necessitate strong semantic parsing capabilities during query understanding and answer generation. The system must accurately identify and handle unit conversions. Multilingual content requires the system to support multiple languages. Additionally, multi-turn conversations must track context to ensure a complete understanding of ADR events and avoid fragmented information.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk Length | 800–1200 characters | Medical texts have strong contextual relevance; longer chunks help retain complete semantics and reduce fragmentation. |
Recall Count | Top 8–12 entries | Pharmacovigilance questions often involve multiple pieces of information; increasing recall count improves relevant recall rate. |
Similarity Threshold | 0.75–0.85 | Ensures recalled documents are highly relevant to the user query, filtering out imprecise or generic information. |
maxContext | 4096 tokens | Multi-turn pharmacovigilance conversations require maintaining a longer context to track complex cases or regulatory details. |
Reranked Return Count | Top 5 entries | Reranks recall results to prioritize the most relevant documents, improving answer quality. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Processing large regulatory documents or multi-page adverse reaction reports requires longer parsing times. |
Common Pitfalls
- Observation: Misunderstanding of medical terminology or unit conversion errors in conversations. Reason: Prompts do not sufficiently guide the model to focus on specialized vocabulary and numerical units, or knowledge base chunking separates specialized terms from their context.
- Observation: The system cannot answer questions about recent regulatory changes or provides outdated information. Reason: The knowledge base update mechanism is incomplete, failing to timely synchronize the latest regulatory documents and guidelines.
- Observation: After a user uploads an adverse reaction report file, the conversation system fails to extract key information or the conversation breaks. Reason: File parsing times out, or the parser cannot correctly identify specific fields in semi-structured medical texts.
Validation Steps
- Conduct multi-turn questioning for typical adverse reaction report cases. Verify if system answers are accurate, complete, and include all key information.
- Upload newly released pharmacovigilance regulatory documents or guidelines. Ask questions about relevant clauses. Verify if the system can correctly cite and explain them, and check the knowledge base update timestamp.
- Select simulated queries containing complex medical terminology, various units of measurement, and multilingual segments. Check the system's ability to understand and process these elements. Evaluate the professionalism and accuracy of the answers.
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.