Multi-turn Conversations and Prompts for Medical Information

Data Characteristics for This Category

Standard answer library data primarily comes from medical information documents published by Marketing Authorization Holders (MAH), product inserts, clinical study reports, adverse event monitoring data, and strictly audited internal medical Q&A sets. The update frequency of this data is relatively stable, typically occurring when drug batches are updated, package inserts are revised, new indications are approved, or significant safety information is released. Document structures are highly standardized, often in PDF, Word, or structured database entries, containing clear titles, paragraphs, charts, and references. Core fields include drug name, active ingredient, indications, dosage and administration, contraindications, adverse reactions, drug interactions, pharmacology and toxicology, clinical trial data, corresponding evidence levels, and publication dates. Units frequently involve measurements (e.g., mg, mL, IU), time (e.g., days, weeks, months), and frequency (e.g., times/day).

Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts

The standardized data structure of standard answer libraries allows for more precise understanding of user questions and intent recognition in multi-turn conversations, reducing ambiguity. The stable update frequency means that knowledge base recall results have high timeliness and accuracy. However, this also requires rapid synchronization of retrieval indexes when knowledge is updated. The rigor of the documents demands that the model's generated responses strictly adhere to the original text, avoiding interpretation or over-generalization. The clarity of fields and units enables precise matching and calculation for numerical questions such as dosage and frequency, enhancing the professionalism of responses. Additionally, multi-turn conversations require sensitivity to context, ensuring that subsequent questions can build coherently on previous answers. The system must also be able to guide users to provide more specific medical information when necessary, narrowing the retrieval scope.

Configuration Settings

Configuration Item	Recommended Value	Rationale for Recommendation
`maxContext`	`8000` tokens	Ensures a sufficiently long conversation context to cover multi-turn medical inquiries and clarifications, reducing forgetfulness.
`Chunk size` (Chunk Length)	`500` characters	Balances the integrity of medical text semantic blocks with retrieval efficiency, avoiding excessive fragmentation or information redundancy.
`Recall count` (Recall Count)	`5` entries	Provides enough candidate information for the model to synthesize, while maintaining recall relevance.
`Similarity threshold` (Similarity Threshold)	`0.78`	Balances recall precision and recall rate, reducing interference from irrelevant information and ensuring the accuracy of medical information.
`Rerank result count` (Rerank Return Count)	`3` entries	After secondary sorting, further filters for the most relevant, high-quality medical response segments.
`Prompt Template` (Prompt Template)	Includes "Based on Provided Medical Information" (Based on the provided medical information)	Forces the model to respond based on recalled content, preventing hallucinations or inaccurate medical advice.

Three Common Mistakes

The message "No Relevant Information Found" (Could not find relevant information) appears in the conversation. This occurs because the Similarity threshold (Similarity Threshold) is set too high, preventing relevant content from being recalled even if it exists.
When users ask about drug dosage or administration, the model provides vague or inaccurate answers. This happens because the Chunk size (Chunk Length) is too small, causing critical numerical information to be split across different segments, leading to incomplete context.
The response.human field returned by the API call is null. This is because the prompt did not explicitly instruct the model to generate a human-readable response, or the model encountered an internal error during generation.

How to Verify Configuration

Conduct multi-turn medical Q&A tests to observe if the model can respond coherently based on context and accurately understand medical terminology.
Randomly select a batch of user questions and check if the knowledge points cited in the model's responses are highly consistent with the original content in the standard answer library.
Simulate users asking numerical questions about specific drug dosages or adverse reactions, and verify that the numbers and units in the model's output are accurate.

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.