Data Characteristics for this Category
Quality documents for autoimmune diseases primarily include clinical trial protocols, investigator brochures, case report forms (CRFs), Good Manufacturing Practice (GMP) related documents, and New Drug Application (NDA)/Biologics License Application (BLA) submissions. Data sources are diverse, encompassing research institutions, pharmaceutical companies, clinical centers, and regulatory bodies. The update frequency is relatively low, typically coinciding with clinical trial phase advancements or changes in regulatory requirements, such as protocol revisions after Phase II clinical trials. Document structures are complex, often containing figures, attachments, and extensive specialized terminology, such as immunosuppressants, biologics, cytokines, and autoantibodies. Fields and units are highly specific, including serological indicators (titer, concentration ng/mL), clinical scoring scales (e.g., SLEDAI, DAS28 scores), adverse event codes (MedDRA terms), and statistical parameters (p-value, 95% CI).
Constraints Imposed by these Characteristics on Multiturn Conversations and Prompts
The specialized and complex nature of autoimmune disease quality documents requires a multiturn conversation system with deep domain understanding. The low update frequency of documents means that knowledge base construction must prioritize managing and retrieving historical versions to ensure conversations can trace back to specific document versions while avoiding outdated information. Figures and attachments within documents pose challenges for information extraction and knowledge representation, requiring consideration of how to effectively integrate non-textual information into the RAG retrieval process. The high specificity of fields and units necessitates precise prompt design to guide the model in understanding and answering questions involving specific indicators and units, for example, distinguishing the clinical significance of different types of autoantibody test results. Tracking conversation logs is crucial for identifying misunderstandings or ambiguities in specialized terminology in user queries, and for optimizing prompts and knowledge retrieval strategies accordingly.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 4096 tokens | Accommodates long contexts in professional documents, reducing truncation risk |
Chunk size (Chunk Size) | 800–1000 characters (characters) | Balances semantic completeness and retrieval efficiency, preventing loss of information in long passages |
Recall count (Retrieval Count) | 8–12 entries (items) | Increases the likelihood of retrieving relevant professional information, covering more details |
Similarity threshold (Similarity Threshold) | 0.75–0.80 | Improves the precision of retrieval results, filtering out irrelevant content |
Rerank result count (Rerank Count) | 5 entries (items) | Selects the most relevant snippets, optimizing final answer generation quality |
CHAT_FILE_EXPIRE_TIME | 0 | Ensures long-term availability of historical quality documents, preventing file expiration |
Three Common Pitfalls
- Inaccurate or missing answers for conversations involving extensive specialized terminology. This occurs because prompts do not adequately guide the model to understand and associate specific domain vocabulary.
- Some files uploaded by users are not indexed during batch processing. This occurs because of complex file formats or file sizes exceeding the
UPLOAD_FILE_MAX_SIZElimit. - The model's understanding of context deviates during multiturn conversations, leading to answers unrelated to previous conversation topics. This occurs because
maxContextis set too low to accommodate the complete conversation history.
How to Verify Configuration
- Conduct multiturn conversation tests, asking specialized questions involving different document versions. Check if the model accurately cites sources and provides consistent answers.
- Upload autoimmune quality documents containing figures and special formats. Verify that the knowledge base successfully extracts key information and that it is retrievable.
- Simulate user questions involving specialized terminology, units, and clinical scoring scales. Evaluate the model's understanding of this specific information and the accuracy of its answers.
- Review conversation logs. Analyze the match between user queries and model retrieval results, and adjust parameters based on the
Similarity threshold(similarity threshold).
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.