Multi-Turn Conversations and Prompts for Metabolic and

Data Characteristics in This Category

Clinical trial data in the metabolic and endocrine disease field primarily comes from Electronic Health Records (EHR), laboratory test reports, imaging reports, and patient self-reported questionnaires. This data updates frequently. Key indicators like blood glucose, blood pressure, and blood lipids may update hourly or daily. Document structures vary. EHRs include unstructured physician handwritten notes, semi-structured admission and discharge records, and structured lab result tables. Specific fields include HbA1c (glycated hemoglobin), FBG (fasting blood glucose), and HOMA-IR (insulin resistance index). Units include mmol/L, mg/dL, and %. Different medical institutions may use different units.

Constraints Imposed by These Characteristics on "Multi-Turn Conversations and Prompts"

High data update frequency requires the knowledge base to quickly synchronize the latest information. This prevents inaccurate pre-screening results due to outdated data. Diverse document structures, especially the presence of extensive unstructured text, challenge information extraction accuracy. Multi-turn conversations must identify critical disease diagnoses, medication history, and comorbidity information from complex medical record descriptions. Field and unit diversity requires prompt design to consider unit conversion and standardization. This ensures the model correctly interprets numerical conditions. For example, when determining if a patient meets an HbA1c < 7.0% inclusion criterion, the model must accurately identify and compare values. This avoids misjudgment due to unit confusion. Additionally, vague statements in patient self-reported information require multi-turn conversations to clarify and guide.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`3000 Tokens`	Covers basic patient information, primary diagnosis, medication history, and current conversation turns. This prevents loss of critical information.
`Chunk size` (Segment Length)	`500 characters` (characters)	Balances textual semantic completeness and retrieval efficiency. Suitable for longer paragraphs in medical records.
`Recall count` (Retrieval Count)	`Top 8 entries` (top 8)	Ensures coverage of various aspects of patient history and test results. Improves hit rate of relevant information.
`Similarity threshold` (Similarity Threshold)	`0.78`	Slightly higher than general thresholds due to the precision requirements of medical terminology. Reduces interference from irrelevant information.
`Rerank result count` (Reranked Return Count)	`Top 3 entries` (top 3)	Further selects the most relevant segments based on high-similarity retrieval. Optimizes context quality.
`System Prompt`	Include unit conversion guidance	Explicitly instructs the model to verify and unify units when processing numerical values like blood glucose and blood pressure.

Three Common Mistakes

The model repeatedly asks for already provided information during a conversation. This results in conversational redundancy. The cause is a maxContext setting that is too small, leading the model to lose early conversation context.
The model fails to correctly identify patient diagnoses or test results from medical records. This results in incorrect pre-screening conclusions. The cause is an improper knowledge base segmentation strategy where critical information is split, or prompts do not explicitly instruct extraction of specific fields.
Debugging preview works correctly, but actual conversations show the knowledge base is not selected. This results in the error message Knowledge base not selected. The cause is usually a mismatch between the deployment environment's knowledge base binding configuration and the debugging environment, or the knowledge base is not published.

How to Confirm Correct Configuration

Select typical virtual medical records of metabolic and endocrine disease patients. Conduct multi-turn conversation tests. Check if the model accurately extracts key indicators such as fasting blood glucose, glycated hemoglobin, and BMI, and records their values.
Simulate patient questions. Verify if the model correctly understands and handles blood glucose unit conversions between mmol/L and mg/dL during the conversation.
Check knowledge base retrieval logs. Confirm that for specific queries, the returned document segments contain core information such as patient medication history and comorbidities. Evaluate the completeness and relevance of the retrieved content.

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.