Data Characteristics in this Domain
Dermatology pharmacovigilance data primarily originates from clinical trial reports, real-world evidence (RWE) studies, physician consultation records, patient spontaneous reports, and drug package inserts. Data update frequencies vary; clinical trial data typically publishes after study completion, while patient spontaneous reports are real-time. Document structures are diverse, including unstructured free text (e.g., handwritten physician notes, patient descriptions), semi-structured tabular data (e.g., adverse event report forms like CIOMS I, MedWatch forms), and structured coded data (e.g., ICD-10 diagnostic codes, MedDRA adverse drug reaction terms). Field specificity is high, often involving specialized dermatological terminology such as lesion morphology (maculopapular rash, vesicles, erythema), location (face, trunk, limbs), severity (mild, moderate, severe), and time course. Units frequently include dosage units (mg/kg), time units (days, weeks, months), and area units (cm²).
Constraints Imposed by These Characteristics on "Multi-turn Conversations and Prompts"
The large volume of free text and specialized terminology in dermatology data requires multi-turn dialogue systems to possess strong semantic understanding and entity recognition capabilities. The real-time and unstructured nature of patient spontaneous reports challenges the system's ability to process high-noise, non-standard language. Multi-turn conversations need to guide users to progressively clarify ambiguous descriptions. For example, when a user mentions "skin redness," the system should be able to ask follow-up questions about the location of redness, accompanying symptoms, and medication history to accurately determine the type of adverse reaction. The existence of standard terminologies like MedDRA means prompt design must balance natural language with professional coding mapping to ensure information accuracy. Diverse document structures imply that the system cannot rely on a single paradigm for information extraction but should integrate multiple information extraction strategies. Ultimately, the system must construct a complete adverse drug reaction event chain from multi-turn interactions to support subsequent evaluation.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8192 tokens | Accommodates complex dermatological medical histories and polypharmacy, ensuring complete dialogue context. |
Chunk size | 500 characters | Balances semantic integrity of long texts with retrieval efficiency, suitable for medical records and reports. |
Recall count | Top 10 entries | Increases coverage of relevant knowledge, reducing issues caused by missing key information. |
Similarity threshold | 0.78 | Ensures retrieved results are highly relevant to dermatological professional terms and symptom descriptions. |
Rerank result count | 5 entries | Prioritizes knowledge most relevant to the current dialogue intent, reducing irrelevant information interference. |
maxResponseTokens | 1024 tokens | Allows generation of detailed explanations and follow-up questions, adapting to the complexity of dermatological diagnosis. |
Three Common Pitfalls
- Dialogue gets stuck at a specific node with no response, and logs show "knowledge base search no results." This may occur if dermatological specialized terminology is not effectively indexed, leading to similarity matching failures.
- During multi-turn conversations, the system frequently re-asks for already provided information, leading to a poor user experience. This happens when the
maxContextparameter is set too low, preventing the system from retaining sufficient historical dialogue context. - API calls to multimodal models result in an "incorrect image format" error. This usually means the input image is not a standard format supported by the model, such as
image/jpegorimage/png.
How to Verify Configuration
- Conduct multi-turn simulated dialogues to verify the system's ability to accurately understand and ask follow-up questions about common dermatological adverse reactions (e.g., rash, urticaria), including location, morphology, accompanying symptoms, and medication history.
- Test the system's effectiveness in extracting key entities from unstructured patient self-reports and matching them with MedDRA terms. Evaluate this by checking the accuracy of generated tags or codes.
- Examine knowledge base retrieval results. After inputting dermatological symptom descriptions, ensure that the retrieved documents or snippets contain relevant content from authoritative drug package inserts, clinical guidelines, or professional literature. Cross-check the reasonableness of the
Similarity threshold.
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.