Data Characteristics
Autoimmune disease pharmacovigilance data originates from clinical trial reports, real-world evidence (RWE) studies, individual case safety reports (ICSRs), and regulatory adverse event databases (e.g., FDA FAERS, EMA EudraVigilance). Data updates frequently, especially after new drug launches or new adverse event reports. Document structures are complex, containing unstructured free text (e.g., patient history, adverse event descriptions, medication details) and semi-structured or structured fields (e.g., MedDRA codes, generic drug names, dosages, routes of administration, adverse event dates, outcomes). Fields often include medical terminology, abbreviations, and multi-language expressions. Units are diverse, covering dosage (milligrams, grams, units), frequency (times/day, weekly), and time (days, months, years).
Constraints Imposed on Multi-turn Conversations and Prompts
The high update frequency of autoimmune disease pharmacovigilance data requires the knowledge base to support rapid synchronization and incremental updates. This ensures multi-turn conversations are based on the latest information. Complex document structures and medical terminology demand highly specialized prompt design to accurately interpret medical concepts and context in user queries. The presence of free text requires the model to have strong information extraction and summarization capabilities in multi-turn conversations, converting unstructured descriptions into structured features suitable for RAG retrieval. Diverse units and numerical expressions challenge the model's understanding and comparison of dosage and frequency information. Prompts must explicitly instruct the model to focus on numerical units and perform appropriate unit conversions or standardization. Additionally, cross-language medical terminology may necessitate multi-language support or specialized terminology mapping.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 6 | Autoimmune pharmacovigilance dialogues often involve tracing multiple adverse event details. |
Chunk size (Segment Length) | 800–1200 characters | Ensures each segment can contain a complete adverse event description or medical concept. |
Recall count (Recall Count) | 10 | Increases the chance of recalling relevant adverse event cases and drug information. |
Similarity threshold (Similarity Threshold) | 0.78 | Balances recall rate and accuracy, reducing interference from irrelevant information. |
Rerank result count (Reranked Return Count) | 5 | Further refines results, focusing on the most relevant medical evidence. |
promptTemplate | Calibrated by testing | Must include guidance for key information such as MedDRA codes, drug information, and dosages. |
Common Pitfalls
- Model response quality degrades after multi-turn conversations, resulting in missing or inaccurate information. This stems from improper context management, where useful information from early turns is truncated or diluted, preventing the model from maintaining a coherent understanding of complex medical histories or multiple adverse events.
- The model fails to effectively extract key information from user-inputted Markdown formatted messages or ignores specific hidden requirements from the user in its replies. This occurs because the prompt does not explicitly instruct the model to process or ignore specific formatting tags, leading the model to parse the format itself as content.
- SQL queries succeed, but results do not display correctly in the AI conversation box. This is due to a mismatch between the SQL query result output format and the input format expected by the conversational model in the workflow or integration configuration, or a missing intermediate step to convert structured data into natural language descriptions.
Verification Steps
- Perform multi-turn conversation tests. Verify the model's ability to accurately reference key information from previous conversations in complex scenarios, such as tracing patient medication history and adverse reaction progression.
- Simulate various queries containing medical terminology, abbreviations, and different units. Check if the model can correctly understand and recall relevant documents from the knowledge base. Verify consistency of key fields and units in the recalled documents.
- Execute a series of test cases with Markdown tags. Verify if the model processes or hides specific formatted content as expected when receiving and sending messages, especially in scenarios involving sensitive information or internal notes.
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.