Multi-turn Conversations and Prompts for Smart Triage

Data Characteristics in This Category

Smart triage data primarily consists of quality documentation. This documentation covers medical service processes, disease diagnosis standards, medication guidelines, and examination items. Sources typically include internal hospital regulations, clinical pathways from national health commissions, drug instruction databases, medical textbooks, and expert consensus. Update frequency is relatively fixed, usually tied to policy changes, medical advancements, or annual revisions. Document structures are complex, containing numerous specialized terms, abbreviations, diagrams, and cross-references. Fields and units are highly specific to medicine, such as disease codes (ICD-10), drug dosages (mg/kg), and examination result units (mmol/L). Precision in values and rigor in descriptions are extremely critical.

Constraints on Multi-turn Conversations and Prompts

The specialized and rigorous nature of medical quality documentation requires multi-turn dialogue systems to precisely identify medical terminology and context when understanding user intent. For example, when a user mentions symptoms, the system must differentiate between chief complaints and accompanying symptoms, then infer based on past medical history. Complex document structures necessitate advanced document parsing capabilities to accurately extract information from different levels, avoiding information fragmentation. The specificity of fields and units challenges prompt engineering. Prompts must guide the model to adhere strictly to medical norms and data formats when generating responses. For instance, when recommending examination items, the system must clearly state the examination's purpose and scope. Furthermore, the relatively fixed update frequency requires knowledge bases to support version management, ensuring conversations always rely on the latest and most authoritative quality documentation.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8` rounds	Balances user intent continuity with model processing complexity, preventing rapid information decay in early rounds.
`Chunk size` (Segment Length)	`400–600 characters`	Accommodates the dense medical terminology and complex sentence structures typical of medical documents, ensuring semantic completeness.
`Recall count` (Recall Count)	`Top 8`	Increases the coverage of relevant document snippets, addressing user questions that may involve multiple knowledge points.
`Similarity threshold` (Similarity Threshold)	`0.75`	Ensures recalled documents are highly relevant to the user query, filtering out low-quality information from fuzzy matches.
`Rerank result count` (Rerank Return Count)	`Top 3`	Further optimizes sorting based on high-quality recall, prioritizing the most accurate and critical information.
`temperature`	`0.3`	Reduces the randomness of model-generated responses, ensuring rigor and accuracy in answers, meeting medical scenario requirements.

Common Mistakes

Model hallucination in conversations, fabricating non-existent diseases or treatment plans: This occurs when prompts provide insufficient constraints on model generation, allowing the model to generate freely without adequate knowledge base support.
Responses are not fully confined to the knowledge base, resulting in generalized answers: This happens when the Similarity threshold (Similarity Threshold) is set too low, or Rerank result count (Rerank Return Count) is too small, leading to insufficient relevant information for the model to provide a complete answer.
User-uploaded images do not display or are not recognized by the model in the chat interface: This indicates that the platform lacks or has incorrectly integrated image parsing capabilities, preventing the model from processing non-text inputs.

How to Verify Configuration

Test typical disease symptom descriptions to see if multi-turn conversations accurately guide to relevant diagnoses or examination suggestions. Verify that responses originate entirely from the knowledge base.
Simulate a user asking vague or less specialized questions. Observe if the model can elicit sufficient information through follow-up questions to initiate accurate knowledge base retrieval. Evaluate the clarity and directness of these follow-up questions.
Import documents containing specific units of measurement or medical codes. Test if the model can accurately restate or cite this information in its responses. Check the correctness of units and values.

Note: 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.