Multi-turn Conversation and Prompts for Rational Drug Use in

Data Characteristics in This Category

Rational drug use data primarily comes from drug inserts, clinical guidelines, drug interaction databases, adverse event reports, and medical literature. This data updates frequently, especially when new drugs launch, indications expand, or safety information changes. Documents are semi-structured and unstructured. For example, drug inserts typically include sections like indications, dosage and administration, contraindications, and adverse reactions. Fields include drug name, active ingredient, dosage, route of administration, patient characteristics, diagnosis, and treatment plan. Units involve mg, ml, μg/kg, times/day, and often include specific medical terms and abbreviations.

Constraints Imposed by These Characteristics on "Multi-turn Conversation and Prompts"

High-frequency data updates require real-time knowledge base synchronization. This ensures that drug information cited in multi-turn conversations is current, preventing inappropriate advice based on outdated data. Semi-structured and unstructured documents make traditional keyword matching difficult for precise retrieval. This necessitates stronger semantic understanding to process complex medical descriptions. Diverse fields and units require prompt design that guides the model to identify and correctly parse information. For example, dosage and route of administration often require inference based on specific patient conditions. In clinical trial pre-screening, conversations often involve multiple dimensions of patient information, such as medical history, medication history, and genetic test results. This leads to longer conversation contexts. The model needs long-context processing capabilities and effective management of multi-turn conversation states to prevent information loss or confusion.

Configuration Strategy

Configuration Item	Recommended Value	Rationale for this Value
`maxContext`	`102400 tokens`	Accommodates patient history, multi-turn Q&A, and reference materials, ensuring complete context.
`Recall count` (Number of Retrieved Items)	`Top 10 entries` (Top 10)	Increases coverage of relevant documents in complex queries, enhancing information comprehensiveness.
`Similarity threshold` (Similarity Threshold)	`0.75`	Balances recall and precision, filtering out low-relevance medical literature or drug information.
`Chunk size` (Segment Length)	`500 characters` (500 characters)	Accommodates longer paragraphs in medical literature, reducing semantic fragmentation and preserving context.
`Rerank result count` (Number of Re-ranked Items)	`Top 5 entries` (Top 5)	Further filters retrieval results, prioritizing the most relevant drugs or plans for the current conversation.
`Model Temperature`	`0.3`	Ensures high certainty and authority in the model's rational drug use recommendations.

Three Common Mistakes

Incorrect drug dosage or usage appears in the conversation. This occurs because knowledge base synchronization is not timely, and the model cites an old version of the drug insert.
After multiple turns, the model fails to accurately associate a patient's previously mentioned allergy history. This leads to recommended drugs conflicting with contraindications. This happens because maxContext is set too low, truncating early conversation information.
The same prompt yields worse results in FastGPT compared to direct use on a large model platform. This is due to an unreasonable knowledge base data segmentation strategy or too few retrieved items, failing to provide sufficient high-quality contextual information.

How to Verify Correct Configuration

Simulate various patient cases. Check if the model can accurately provide rational drug use recommendations after multi-turn conversations. Verify the consistency of recommendations with the latest drug inserts.
Review maxContext actual usage in conversation logs. Confirm if context truncation occurs, especially in long conversation scenarios.
Use FastGPT's debugging interface. Check the content and similarity scores of each retrieved document. Evaluate the relevance of retrieval results and adjust Similarity threshold (Similarity Threshold) and Recall count (Number of Retrieved Items).
Cross-reference key fields like drug name, dosage, and usage in the model's output. Ensure they completely match factual data in the knowledge base.

Note: The values provided are common starting points. Measure them against specific samples to determine optimal settings.

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.