Multi-Turn Conversations and Prompts for Smart Triage

Data Characteristics

Smart triage data primarily comes from medical knowledge bases, drug inserts, disease diagnostic criteria, clinical guidelines, and user medical records. This data updates frequently. New drug approvals, treatment plan adjustments, or policy changes require immediate synchronization. Document structures typically include both structured and unstructured content. Examples include disease symptoms, causes, diagnoses, treatment plans; drug ingredients, indications, dosages, contraindications; and patient medical history descriptions and examination reports. Fields and units are highly standardized in medical terminology, covering dosages (mg, g), frequencies (times/day), and durations (days, weeks).

Constraints on Multi-Turn Conversations and Prompts

High-frequency medical data updates require a flexible knowledge base synchronization mechanism. This ensures the timeliness of information cited in multi-turn conversations. The coexistence of structured and unstructured data means prompt design must balance precise extraction with semantic understanding. For example, prompts must identify key symptoms from medical history descriptions and match them with structured disease characteristics. The professional and standardized nature of medical terminology constrains the accuracy of terms used in prompts, preventing ambiguity. Incorporating user medical records requires multi-turn conversations to understand and integrate personal user information for personalized consultations. This also demands strict adherence to data privacy and compliance requirements to prevent sensitive information leakage.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`maxContext`	`2048 token`	Ensures the context covers key symptom descriptions and medical history information across multiple user turns, preventing omissions.
`Recall count` (Recall Count)	`Top 10 entries` (Top 10)	Expands the knowledge base recall scope, increasing the likelihood of identifying potentially relevant diseases and medications, and reducing misdiagnosis risk.
`Similarity threshold` (Similarity Threshold)	`0.75`	Balances recall precision and breadth. This prevents misjudgments due to semantic deviation while retaining some generalization capability.
`Chunk size` (Segment Length)	`400 characters` (400 characters)	Accommodates the long descriptive nature of medical documents. This ensures each segment contains a complete medical concept or treatment step.
`Rerank result count` (Rerank Return Count)	`Top 5 entries` (Top 5)	Filters initial recall results a second time. This prioritizes displaying treatment suggestions that best match the user's current symptoms.
`temperature`	`0.5`	Ensures the model's output is professional and rigorous. This reduces the risk of generating speculative or inaccurate medical advice.

Common Mistakes

Empty operational data in conversation logs: This may occur if data flow configuration is incorrect after invoking an interface workflow. Key variables might not transfer correctly to the logging module.
"Cannot copy" prompt after clicking the copy button, with a pop-up requiring manual copying: This usually happens because browser security policies restrict direct JavaScript access to the clipboard. Alternatively, the FastGPT frontend code might not support the copy operation across all browser environments.
The settings button disappears after changing to "Variable Reference," preventing adjustment of parameters like temperature: This may be because the system design defaults to preset parameters in variable reference mode, without providing a separate parameter adjustment entry.

Verification

Simulate various typical user consultation scenarios, including symptom descriptions, medication inquiries, and disease prevention. Check if the multi-turn conversation flow is smooth and if the model's answers are accurate and medically sound.
Verify the sources of knowledge points cited in conversations. Ensure they originate from the specified medical knowledge base. Check the knowledge base version and update time to confirm data timeliness.
Intentionally introduce vague or ambiguous medical terms in different conversation turns. Observe if the model correctly guides the user for clarification or provides reasonable explanations.
Examine conversation logs. Confirm that parameters like Recall count (Recall Count) and Similarity threshold (Similarity Threshold) for each knowledge base retrieval are effective as expected. Also, confirm maxContext effectively controls context length.

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.