Data Characteristics in this Category
Pharmacovigilance (PV) registration documents primarily include Adverse Drug Reaction (ADR) reports, Periodic Safety Update Reports (PSUR/PBRER), Risk Management Plans (RMP), and Post-Authorization Safety Study (PASS) data. Data sources are diverse, encompassing clinical trial data, real-world data (RWD), literature search results, patient reports, and healthcare institution reports. Updates are frequent, especially during the initial post-marketing period, with ADR reports submitted quarterly or annually, and safety update reports typically on a semi-annual, annual, or triennial cycle. Document structures usually follow ICH E2B or E2C guidelines, often in structured or semi-structured XML, PDF, or Word formats. Fields involve patient information (de-identified), drug information, adverse reaction terms (MedDRA coding), dosage, event occurrence time, and causality assessment results. Units primarily focus on dosage (mg, g, ml, etc.), time (days, hours, years), and frequency (times/day, cases).
Constraints Imposed by these Characteristics on "Multiturn Conversation and Prompts"
The multi-source and complex structure of pharmacovigilance data challenge the accuracy of multiturn conversations. The use of specialized terminology like MedDRA requires the model to possess domain knowledge; otherwise, semantic deviations can occur. Frequent data updates mean the knowledge base needs rapid synchronization to ensure conversations are based on the latest information. For example, after a new PSUR report is published, the model should immediately be able to answer questions about newly added adverse reaction information within it. The ability to extract key information from semi-structured documents directly impacts conversation quality, especially when querying the incidence or causality of specific adverse reactions. Furthermore, the rigor of registration documents demands high traceability for conversation results, capable of pointing to specific sections or report numbers for information sources. Precise understanding of units like dosage and time is critical to avoid misunderstandings and ensure compliant responses.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 2000 characters | Ensures context covers detailed descriptions and relevant assessments of adverse event incidents in registration documents, preventing loss of critical information. |
chunkLength | 400 characters | Balances semantic completeness of text with recall granularity, adapting to itemized or paragraph-style descriptions in adverse reaction reports. |
recallCount | top 8 items | Increases the probability of recalling relevant snippets from vast registration data, improving accuracy for complex queries. |
similarityThreshold | Calibrated by actual measurement | Avoids recalling too much irrelevant information while ensuring capture of specialized terminology and synonyms. |
rerankReturnCount | top 3 items | After reranking, focuses on a few high-quality results most relevant to user intent, enhancing response efficiency. |
systemPrompt | Detailed definition of role and knowledge boundaries | Clarifies the AI's role as a pharmacovigilance document assistant, limits the scope of answers to existing registration documents, and emphasizes compliance and data sources. |
Three Common Mistakes
- Conversation freezes or no response: Often due to
maxContextbeing set too large, causing the model to take too long to process long contexts, or upstream knowledge base retrieval timing out. - AI answers include unmentioned adverse reactions or incorrect dosages: This occurs because knowledge base data synchronization is not timely, the model failed to access the latest pharmacovigilance reports, or extracted field values are inaccurate.
- API calls return null or incomplete results: Typically due to
similarityThresholdbeing set too high, leading to filtering out valid information with slightly lower relevance, orrecallCountbeing too small to cover comprehensive information.
How to Confirm Proper Configuration
- Submit a series of queries targeting specific adverse reactions, dosages, or causal relationships. Cross-reference key information in AI responses with original submission documents to ensure accuracy.
- Simulate submitting queries about new adverse events in the latest PSUR report. Verify if the AI can provide relevant information promptly to confirm the effectiveness of the knowledge base update mechanism.
- Test complex queries containing MedDRA terminology. Check if the AI correctly understands and provides relevant report details, validating the matching of domain knowledge.
- Query via API calls. Check if the completeness of fields and data types in the returned results match expectations, ensuring correct data extraction and transmission.
The values provided are common starting points and should be measured against specific 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.