Data Characteristics
Dermatology quality documents include clinical pathways, treatment guidelines, drug inserts, device manuals, adverse event reports, Standard Operating Procedures (SOPs), and Good Clinical Practice (GCP) related files. These documents have a low update frequency, typically revised quarterly or annually. Drug inserts and device manuals depend on regulatory approvals or manufacturer updates. Document structures are hierarchical, containing chapters, sections, figures, and appendices. Fields often involve disease diagnosis codes (e.g., ICD-10), drug dosage units (mg/kg, IU), device parameters (wavelength, energy density), and biomarker units (ng/mL, U/L).
Constraints Imposed by These Characteristics on Multiturn Conversation and Prompts
The hierarchical structure and specialized fields in dermatology quality documents demand high accuracy in contextual understanding for multiturn conversations. The conversation model must precisely identify and link cross-chapter references. For example, when a user asks about a drug's side effects, the model needs to recall the drug insert and potentially link to specific cases in adverse event reports. Recognizing specialized units and abbreviations (e.g., "QD" for once daily, "BID" for twice daily) is crucial to avoid misinformation due to unit confusion. Low update frequency means the model prioritizes memory and retrieval of historical data. It also needs to differentiate between old and new document versions to ensure the latest, most authoritative guidance is provided.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 20 entries | Dermatology documents are highly interconnected. This ensures effective historical information traceability in multiturn conversations. |
Chunk size | 500 characters | Balances document detail with retrieval efficiency, preventing key information dilution by overly long segments. |
Recall count | Top 8 entries | Increases recall coverage to address professional terminology ambiguity and potential associations. |
Similarity threshold | 0.75 | Improves matching accuracy for scenarios with specialized terminology and high precision requirements. |
Rerank result count | 5 entries | Reranking more accurately filters document segments most relevant to the current turn. |
promptTemplate | Calibrate based on actual testing | Must include clear instructions to guide the model to focus on critical information like disease codes and dosage units. |
Common Pitfalls
- Replies fail to link to context. This often occurs when the
maxContextparameter is set too low, preventing the model from effectively using previous conversation history. - Conversation logs show misinterpretation of some specialized terms. This happens when the prompt does not explicitly instruct the model to focus on or explain specific domain-specific vocabulary and units.
- The same question yields different answers across document versions. This indicates the knowledge base lacks proper version management configuration or retrieval does not prioritize the latest version.
How to Verify Configuration
- Randomly select diagnostic questions for specific dermatological conditions. Conduct multiturn conversation tests to check if replies accurately cite relevant document content and if the cited document version is the latest.
- For questions containing special units (e.g., mg/kg, IU) or abbreviations (e.g., QD, BID), check if the model correctly understands and provides standard-compliant answers.
- Verify if the model can consistently link and deepen its answers in subsequent turns when the user repeatedly mentions a symptom or drug in the conversation. This can be assessed by observing the actual utilization of
maxContext.
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.