Data Characteristics
Dermatology regulation data originates from the National Health Commission, various medical institutions' rules and regulations, drug inserts, treatment guidelines, and academic journals. Data updates are relatively stable, typically revised periodically after regulatory policy releases. Drug inserts and treatment guidelines may update with new drug approvals or clinical research advancements. Overall update cycles are long, often quarterly or semi-annually. Document structures are primarily PDF, Word, and Markdown formats. Content includes legal provisions, technical operation specifications, medical record management requirements, drug contraindications, and common dermatological treatment pathways. Fields cover disease names, diagnostic criteria, treatment plans, drug dosages, operating procedures, and adverse reactions. Units are mostly international standard units (e.g., mg, ml, μg/kg), time units (e.g., hours, days), and percentages.
Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall
The stability and low update frequency of dermatology regulations mean significant data cleaning and preprocessing efforts during initial knowledge base construction, but subsequent maintenance costs are relatively low. Complex document structures, containing numerous specialized terms, nested article numbers, and tabular data, challenge text segmentation and semantic understanding. For example, conditional branches in treatment pathways and drug dosage ranges require retrieval results to accurately point to specific regulations for specific scenarios. Accurate identification of specialized fields and units directly impacts answer precision. Recall must specifically focus on whether this information is correctly extracted and matched. Long documents may contain multiple synonymous but differently phrased contents, requiring enhanced generalization capabilities for recall. Furthermore, regulatory Q&A demands extremely high authority and accuracy in recall; any deviation can lead to misunderstandings.
Configuration Settings
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
Chunk size (Segment Length) | 800–1200 characters (characters) | Dermatology regulation documents are typically long and contain multi-level logic. Increasing segment length helps retain contextual completeness and reduces semantic fragmentation. |
Chunk Overlap Length (Segment Overlap Length) | 100–200 characters (characters) | Ensures sufficient semantic overlap between adjacent text blocks, preventing loss of important information at segmentation boundaries and improving retrieval recall rate. |
Recall count (Recall Count) | 5–8 entries (items) | Given the precision requirements of regulatory Q&A, recalling more relevant items helps the model understand and integrate information from multiple perspectives, improving the comprehensiveness of the final answer. |
Similarity threshold (Similarity Threshold) | Calibrate by actual measurement | Adjust the threshold based on actual test results to balance recall rate and precision, ensuring recalled content is highly relevant to the user query. |
Rerank result count (Rerank Return Count) | 3 entries (items) | After initial recall, use a reranking mechanism to further filter the three most relevant items, reducing the processing burden on the subsequent language model and improving response efficiency. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Processing large PDF or Word format regulation documents can be time-consuming. Increasing the timeout prevents parsing interruptions. |
Three Common Mistakes
- Query results contain incomplete regulatory provisions or operating procedures. This happens when text segmentation breaks the integrity of a provision, causing individual text blocks to fail in expressing complete semantics independently.
- Answers cite drug dosages or treatment plans that do not match actual regulations. This occurs when the knowledge base does not standardize unit-bearing numerical values, leading to retrieval matching deviations.
- When facing complex queries, the number of recalled items is too small or their relevance is insufficient. This results from setting
Similarity threshold(Similarity Threshold) too high, filtering out potentially relevant document blocks with slightly different phrasing.
How to Confirm Proper Configuration
- Select a batch of representative dermatology regulation questions and answers. Check if the recall results include all necessary information points and can point to specific locations in the original documents.
- For queries involving critical information like drug dosages or operating procedures, verify that the numerical values and units in the recalled content are accurate and consistent with original regulatory provisions.
- Simulate user queries containing synonyms or different phrasing. Observe if recall results can still accurately match relevant regulatory provisions.
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.