Multi-Turn Conversations and Prompts for Metabolic and

Data Characteristics

Metabolic and endocrine protocols and SOP documents originate from internal management guidelines, clinical diagnostic guidelines, drug inserts, and equipment operation manuals. These come from hospitals, pharmaceutical companies, and medical device manufacturers. Document updates are relatively stable, typically occurring every few months to a year. Revisions happen immediately with regulatory changes or new drug launches.

Document structures feature clear hierarchical chapters and clauses. They often include many tables, diagrams, and flowcharts. Content covers disease diagnostic criteria, treatment plans, drug dosages, monitoring indicators, and adverse reaction management. Common fields include "medication dosage (unit: mg/kg/day)", "blood glucose monitoring frequency (unit: times/day)", and "insulin type (unit: U)". Accuracy for values and units is critical.

Constraints on Multi-Turn Conversations and Prompts

The data characteristics of metabolic and endocrine protocol documents impose specific requirements on multi-turn conversation and prompt design.

First, the document update cycle dictates the knowledge base synchronization strategy. Regular full or incremental updates are necessary to ensure conversation accuracy. Second, complex hierarchical structures and numerous tables/diagrams mean chunking strategies must consider semantic completeness to avoid truncating critical information.

In multi-turn conversations, users may frequently mention specific drug dosages or monitoring frequencies with units. The model must accurately understand and extract relevant numerical information from documents, performing unit conversions or comparisons.

Prompt design needs to guide the model to focus on key diagnostic criteria and treatment processes, preventing deviation from the core topic in multi-turn conversations. It must also handle user inquiries about sensitive information like adverse reactions and contraindications.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Length)	`500–800 characters`	Ensures each chunk contains a complete medical concept or operational step, aiding context understanding.
`Recall count` (Recall Count)	`8–12 items`	Considering the complexity of disease diagnosis and treatment, increasing recall covers more relevant clauses.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Guarantees precision of recalled content, filtering out paragraphs not strongly related to medical concepts.
`Rerank result count` (Rerank Return Count)	`4–6 items`	Streamlines key information presented to the user while maintaining accuracy.
`maxContext`	`4000 tokens`	Allows the model to retain sufficient historical information in multi-turn conversations for complex logical reasoning and cross-referencing.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Handles large PDFs or SOP files with complex tables, preventing parsing timeouts.

Common Mistakes

Responses like "cannot find relevant dosage information" or "unit mismatch" in conversations. This usually happens when document parsing fails to correctly identify numerical values with units, or prompts do not guide the model to perform unit matching.
The model "hallucinates" or repeatedly mentions irrelevant content in later stages of multi-turn conversations. This might be due to maxContext being too small, leading to truncated conversation history and loss of context.
After a user uploads a new SOP file, conversation content still relies on the old version. This occurs when the knowledge base synchronization strategy is not configured for automatic triggering, or file parsing failures prevent new data from being ingested.

Verification

Select typical multi-turn Q&A scenarios in the metabolic and endocrine domain, such as "Insulin Dosage Adjustment Process for Diabetic Patients". Verify if the model's responses in different turns match document content, focusing on numerical and unit accuracy.
Test Q&A involving complex document structures like tables and flowcharts. Observe if the model can correctly extract and explain key information, such as threshold values for diagnostic indicators in "Diagnostic Criteria for Hyperthyroidism".
Test with different versions of the same protocol document. Verify that after knowledge base updates, the model answers based on the latest version, and old version information is no longer cited.
Check file parsing status in system logs. Ensure all uploaded documents are successfully parsed and ingested, without PARSE_FILE_TIMEOUT_SECONDS or parsing errors.

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.