Multi-turn Conversation and Prompts for Mental Illness

Data Characteristics for this Category

Registration and declaration data for mental illnesses come from various sources. These typically include clinical trial reports (Phases I, II, III), non-clinical study reports, pharmaceutical research data, epidemiological data, real-world patient data (RWD), and regulatory documents for similar approved drugs. Data updates are relatively infrequent, primarily occurring when different phases of clinical trials release reports and when review bodies provide feedback. Document structures are complex, containing extensive specialized terminology, medical abbreviations, and charts. Common document formats include PDF, Word, and Excel. Field and unit standardization requirements are high. Examples include dosage units (mg, μg), time units (days, weeks, months), statistical indicators (p-values, confidence intervals), and scale scores (e.g., HAM-D, PANSS scores). Differences in regulatory standards across various countries and regions are also common.

Constraints from these Characteristics on Multi-turn Conversations and Prompts

The complex data characteristics of mental illness registration documents impose specific constraints on multi-turn conversation and prompt design. The specialized and multi-source nature of the documents requires prompts to accurately understand medical terminology and contextual relationships, avoiding information bias due to ambiguity. Infrequent data updates mean that knowledge base construction must prioritize historical version management, ensuring conversations always rely on the latest or specified version of the declaration documents. The presence of complex document structures and charts increases the difficulty of information extraction. Prompt design needs to consider how to guide the model to process non-textual information or its descriptions. The strict standardization of fields and units requires the conversational system to accurately cite or convert this information when generating responses. This is especially critical when comparing different research data, where aligning scales and indicators is essential to avoid misunderstandings due to inconsistent units.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`maxContext`	`8`	Ensures coverage of a sufficiently long historical context in multi-turn conversations to handle complex medical logic chains.
`Chunk size` (Chunk Length)	`800–1200 characters` (characters)	Balances semantic integrity for lengthy medical reports with model processing efficiency, preventing information loss from overly long texts.
`Recall count` (Recall Count)	`Top 5–8 entries` (top 5–8)	Given the depth and interconnectedness of specialized data, increasing the recall count helps cover more relevant but not directly matching information snippets.
`Similarity threshold` (Similarity Threshold)	`0.75`	Given the specialized and precise nature of mental illness terminology, a higher threshold ensures strong relevance of recalled content.
`Rerank result count` (Reranked Return Count)	`3`	Further filters the most core and relevant 3 pieces of information from the recalled set through reranking, improving the quality of the final response.
`Temperature`	`0.3`	Reduces the randomness of model-generated responses, ensuring the rigor and accuracy of output content in medical declaration scenarios.

Three Common Pitfalls

Workflow conversation fails, but the model backend receives response logs: This typically indicates that a downstream node in the workflow configuration failed to correctly parse or process the data format returned by an upstream node, leading to process interruption.
Historical conversation not passed when calling searchTest: This suggests that when integrating or calling the RAG module, the chat_history parameter was not correctly mapped or passed, preventing the model from utilizing context for retrieval augmentation.
Application conversation shows an error, but no logs are found: This could be due to a high log level setting or an incorrect log configuration path, preventing the system from recording detailed error stack information.

How to Verify Correct Configuration

Ask multi-turn questions on key medical terms. Observe if the model can consistently and accurately extract and integrate information from different documents. This helps determine if Similarity threshold (Similarity Threshold) and Recall count (Recall Count) are appropriate.
Use documents containing chart or table descriptions for Q&A. Check if the model can understand and cite their content. This evaluates the effectiveness of the chunking strategy for handling non-textual information.
Simulate complex logical reasoning scenarios from declaration documents, such as comparing clinical efficacy data of two drugs. Examine if the model maintains consistency and logical coherence in multi-turn conversations. This validates the effectiveness of maxContext.

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.