Data Characteristics for This Category
Core data for dermatology registration and declaration documents originates from clinical trial reports, non-clinical study reports, pharmaceutical research data, and literature reviews. Clinical data includes patient inclusion criteria, disease diagnostic standards, efficacy evaluation indicators (e.g., Psoriasis Area and Severity Index PASI, Eczema Area and Severity Index EASI), and adverse event incidence rates. Data updates typically occur at different stages of clinical trials or as per regulatory requirements. Document structures are complex, often presented as Clinical Study Reports in ICH E3 format or CTD modular documents. Fields involve extensive medical terminology, units of measurement (e.g., mg/kg, %, cm²), and various scoring scales. Data often mixes structured tables, semi-structured text, and unstructured charts.
Constraints Imposed by These Characteristics on "Multi-turn Conversations and Prompts"
Dermatology declaration documents are highly specialized medically. This requires the multi-turn conversation system to accurately understand complex medical terminology and disease mechanisms. The precision of measurement units and scoring scales means prompts must emphasize strict matching of values and units, avoiding vague interpretations. The mix of multimodal data (text, tables, charts) demands higher document parsing capabilities to ensure RAG retrieval covers all key information. Iterative updates to clinical trial reports mean the knowledge base needs to support efficient version management and incremental updates to ensure the timeliness of conversation content. Additionally, varying regulatory requirements across countries and regions necessitate prompts that can differentiate and handle regional guidelines, such as specific FDA, EMA, or NMPA guidelines.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 600–800 characters (characters) | Preserves contextual integrity while balancing retrieval efficiency and semantic relevance. |
Recall count (Recall Count) | Top 8–12 entries (top 8–12 entries) | Addresses the need to cover complex medical concepts and multi-dimensional evidence. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Ensures retrieved results are highly relevant to specialized dermatology queries. |
maxContext | 4000 token | Supports understanding lengthy medical background information in multi-turn conversations. |
Rerank result count (Reranked Return Count) | Top 5 entries (top 5 entries) | Filters for the most relevant core evidence, reducing interference from irrelevant information. |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds (seconds) | Accommodates parsing time for large clinical study reports or pharmaceutical data. |
Three Common Mistakes
- Observation: The AI fails to extract
PASIscore trend changes from uploaded documents, consistently responding "no relevant information found." Reason: Prompts do not explicitly guide the AI to focus on charts or specific tabular data within the document, instead focusing only on text content retrieval. - Observation: In a multi-turn conversation, when a user asks about the unit of a certain indicator, the AI's answer is inconsistent or uses generic units. Reason: Medical measurement units were not standardized during knowledge base data import, leading to varied unit representations in retrieval results.
- Observation: After refreshing a conversation, previous conversation history disappears, showing as a new conversation. Reason: System session management mechanism is improperly configured, failing to correctly persist session states or associate
session IDs.
How to Confirm Proper Configuration
- Conduct multi-turn questioning on a clinical trial report for a specific dermatological disease (e.g., psoriasis). Verify if the numerical values and units of key efficacy indicators (e.g.,
PASIimprovement rate) in the AI's response match the original report. - Upload a non-clinical study report containing complex charts and tables. Ask questions about drug mechanisms of action and toxicology data. Check if the AI can accurately identify and cite data points from the charts.
- Simulate a user repeatedly asking about different aspects of the same concept in a conversation. Observe the AI's ability to remember and reference context in multi-turn dialogues, for example, asking about the correlation between drug dosage, administration route, and adverse reactions.
The values provided are common starting points and should be measured 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.