Data Characteristics
Market access pharmacovigilance data primarily comes from Marketing Authorization Holders (MAH) submissions. These include Risk Management Plans (RMP), Periodic Safety Update Reports (PSUR), Individual Case Safety Reports (ICSR), and regulatory documents like drug labels and warning letters. These documents are typically in PDF, Word, or XML formats, with varying degrees of structural organization. Data updates are frequent; ICSRs might be submitted in real-time or periodically, while RMPs and PSURs have fixed reporting cycles. Documents contain extensive medical terminology, drug names, adverse event (AE) descriptions, patient characteristics, dosage information, and causality assessments. For example, core fields in an ICSR report include SafetyReportId, PatientAge, DrugName, AdverseEventTerm, and Outcome.
Constraints on Multi-Turn Conversations and Prompts
The complexity and specialized nature of market access pharmacovigilance data impose specific requirements on multi-turn conversation and prompt design. First, diverse and frequently updated data sources require FastGPT's knowledge base to efficiently ingest and index various document formats and support incremental updates. This ensures the conversational model accesses the latest risk information. Second, the extensive medical terminology and abbreviations in documents necessitate carefully designed prompts. These prompts guide the model in accurate entity recognition and relationship extraction, preventing misinterpretations or omissions of critical information. For example, when querying AE, the model must differentiate between "Adverse Event" and other meanings. Finally, multi-turn conversations may involve comparative analysis of specific drug safety data across different post-marketing phases. This requires prompts that guide the model to integrate and infer information across multiple related documents. Examples include comparing the incidence rate changes of a specific adverse event across different PSUR cycles or tracing the entire process of a risk signal from early warning to regulatory action.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000 tokens | Accommodates lengthy RMP and PSUR documents, ensuring the model can process sufficient context information in a single pass. |
Chunk size | 500 characters | Balances content completeness and retrieval efficiency. Avoids excessively long segments that lead to information redundancy or overly short segments that cause semantic fragmentation. |
Recall count | 10 items | Ensures enough relevant document snippets are retrieved from the large knowledge base, increasing information coverage. |
Similarity threshold | 0.78 | Balances recall accuracy and breadth. Filters out low-relevance results while retaining highly relevant information. |
Rerank result count | 5 items | Re-ranks initial retrieval results, ensuring the most relevant core information is displayed first to improve model processing efficiency. |
temperature | 0.3 | Ensures the stability and accuracy of the model's output. Reduces the randomness of generative responses, aligning with the rigorous requirements of pharmacovigilance. |
Common Mistakes
- The model fails to correctly understand or associate medical terms in a conversation. This occurs because the knowledge base lacks corresponding glossaries or entity recognition training data.
- When a user asks about historical safety data trends for a specific drug, the model only provides information from a single report. This happens because the knowledge base did not effectively link or version manage similar documents from different time points during ingestion.
- After a user uploads an attachment, the conversation cannot continue, or the attachment content is not recognized. This is due to the
UPLOAD_FILE_MAX_SIZEparameter being set too low or the file parsing plugin not adapting to specific document formats.
How to Verify Configuration
- Select a drug with multiple PSURs. Ask about the incidence rate changes of a specific adverse event across different reporting periods. Check if the model can integrate and compare the data.
- Query the definition of a rare adverse event, related drugs, and regulatory actions taken. Verify the accuracy of the model's professional terminology explanations and information traceability.
- Upload an RMP document containing complex tables and charts. Then, ask about key risks and risk management measures within the document. Check if the model can accurately extract both structured and unstructured information.
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.