Data Characteristics
Gene therapy AAV (adeno-associated virus) pharmacovigilance data originates from clinical trial reports, real-world evidence (RWE) studies, post-market regulatory reports, and academic literature. This data typically exists as unstructured text, such as case reports, follow-up records, and medical journal articles. It also includes structured laboratory results and basic patient information. Data updates frequently; clinical trial data releases periodically as studies progress, and post-market data accumulates continuously. Document structures are complex, often containing extensive medical terminology, gene sequence information, viral vector design details, and adverse event descriptions. Key fields include patient ID, AAV serotype, vector dose, administration route, target gene, adverse event (AE) description, severity, onset time, outcome, and causality assessment. Units involve viral genome copies (vg/mL), dose (vg/kg), and time (days, weeks, months).
Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts
The complexity of gene therapy AAV pharmacovigilance data places specific demands on multi-turn conversation and prompt design. Medical terminology and specialized abbreviations in unstructured text require high-precision entity recognition from the model to avoid misinterpreting critical information. Multi-turn conversations need to track multiple adverse events reported by a patient at different times and link them to AAV vector details for correlation analysis. For example, if a user asks, "Is the elevated liver enzyme in this patient after administration related to the AAV vector dose?", the system must extract patient dose, liver enzyme test results, and AAV serotype information from historical conversations and the knowledge base for a comprehensive assessment. Furthermore, high data update frequency requires the knowledge base to quickly synchronize with the latest research advancements and regulatory dynamics, ensuring the timeliness and accuracy of conversation content. Prompt design must precisely guide user questions, for instance, by explicitly asking about the onset time, severity, or relationship of an adverse event to a specific AAV vector, to narrow the search scope and improve answer quality.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
maxContext | 6 turns | Balances contextual understanding in complex scenarios with conversational efficiency, preventing performance degradation from excessively long contexts. |
Chunk size (Segment Length) | 800–1200 characters | Balances text block completeness and recall efficiency, ensuring a single segment can contain a complete adverse event description or key research conclusion. |
Recall count (Recall Count) | top 5 | Covers the most relevant knowledge points, reduces interference from irrelevant information, and improves retrieval efficiency. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Effectively filters out low-relevance documents, focusing on specialized content in gene therapy AAV pharmacovigilance. |
Rerank result count (Reranked Return Count) | top 3 | Further refines results, ensuring the information presented to the user is the most core and relevant. |
prompt | Calibrated by actual measurement | Dynamically adjusts prompts based on query requirements for specific AAV serotypes or target genes to optimize recall and generation. |
Common Mistakes
- An error message displays "API Key invalid or insufficient permissions." This occurs when an incorrect API Key type is used, such as attempting to call an application conversation interface with a global generic API Key, or if the Key lacks the corresponding application permissions.
- The model fails to accurately associate patient adverse events mentioned in different turns of a conversation. This happens when the
maxContextparameter is set too low, causing the model to lose critical contextual information and preventing effective multi-turn reasoning. - The answer contains a large amount of general medical knowledge unrelated to AAV pharmacovigilance. This is due to a
Similarity threshold(similarity threshold) set too low, which recalls many non-specific document segments, diluting the specialization.
How to Confirm Proper Configuration
- Conduct multi-turn conversation tests to verify if the model accurately understands and associates patient information, AAV vector details, and adverse event descriptions mentioned in previous turns.
- Simulate queries for adverse event incidence or severity of specific AAV vectors, checking if the model can recall and integrate relevant data from the knowledge base.
- Submit complex questions containing specialized medical terminology and gene sequence information, evaluating the accuracy and professionalism of the model's answers, and confirming that the content aligns with the gene therapy AAV pharmacovigilance domain.
Note: The values provided are common starting points. Measure them 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.