Data Characteristics
Dermatology product data comes from diverse sources. These include clinical trial reports, drug inserts, medical journals, pathology image analysis results, and patient feedback. Data update frequencies vary. Drug inserts and clinical guidelines typically update quarterly or annually. Clinical trial data may release in real-time or consolidate in phases.
Document structures also vary. Drug inserts usually contain standardized fields such as indications, dosage and administration, adverse reactions, and contraindications. Clinical trial reports are more complex. They involve subject characteristics, treatment regimens, efficacy endpoints, and safety assessments.
Field examples include ICD-10 disease codes, ATC drug classification codes, CAS chemical substance registration numbers, and NDC national drug codes. Common dosage units are mg, g, and ml. Efficacy metrics may involve PASI scores, percentage of body surface area (% BSA), and visual analog scale (VAS) for itching.
Constraints on Tool Calling and Plugins
The diversity and update frequency of dermatology product data impose specific requirements on tool calling and plugins. Standardized structures, like those in drug inserts, facilitate key information extraction using regular expressions or structured data parsing tools. Unstructured clinical trial reports, however, require more complex Natural Language Processing (NLP) tools for information extraction.
High-frequency data sources, such as new clinical trial results, demand tools capable of periodic or on-demand data synchronization and updates. This ensures the knowledge base remains current. Industry-specific fields, like ICD-10 codes, require precise matching or mapping of professional terminology in tool calls to avoid ambiguity. Standardized dosage units and efficacy metrics necessitate unit conversion or numerical calculations in generated responses. This requires plugins with mathematical computation capabilities or the ability to call external calculation services.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 3000-4000 characters | Accommodates detailed descriptions in dermatology literature while balancing model processing capacity and cost-effectiveness. |
Recall Count | 8-12 items | Ensures coverage of multi-dimensional information for relevant drugs or diseases, avoiding critical detail omissions. |
Similarity Threshold | 0.75-0.85 | Filters out irrelevant document segments, improves retrieval accuracy, and ensures contextual relevance. |
Reranked Return Count | 4-6 items | Focuses on the most relevant core information, reducing the model's burden of processing redundant information. |
Tool Call Timeout | 30 seconds | Accounts for variations in external API response times, ensuring smooth tool call processes. |
API_KEY (External Tool Auth) | Calibrated based on actual measurements (e.g., xxxx-yyyy-zzzz format) | Ensures secure access to external knowledge bases or computational services. |
Common Pitfalls
- Tool calls return empty or incorrectly formatted fields. This occurs when the external API's data structure does not match predefined parsing rules, or when necessary error handling logic is missing.
- Unit conversion errors lead to unexpected results when processing drug dosages or efficacy metrics. This happens due to insufficient consideration of different unit systems used by various data sources or incorrect configuration of conversion functions.
- Model performance degrades within the tool calling workflow, for example, with the DeepSeek R1 model. This can be due to conflicts between the
Knowledge Base Plugin Optimizationfeature and tool calling logic, or the model's instruction following capabilities for specific tools need improvement.
Validation
- Simulate multiple conversational scenarios involving different dermatology product inquiries. Observe if tools trigger correctly and check the accuracy and completeness of their returned results.
- Review tool call logs. Confirm external API request parameters and response data meet expectations. Ensure no
4XXor5XXstatus codes appear. - Compare FastGPT-generated responses with original documents. Verify the accuracy of cited information, especially for critical fields like dosage, indications, and adverse reactions.
- For queries involving numerical calculations or unit conversions, manually verify calculation results. Ensure the plugin's processing logic is correct.
The values provided are common starting points and should be measured against specific 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.