Data Characteristics for This Category
Rational drug use product data originates from drug inserts published by the National Medical Products Administration (NMPA), clinical guidelines, drug interaction databases, adverse reaction reports, and pharmacological literature. This data updates frequently. Drug inserts and clinical guidelines, in particular, may revise periodically due to new drug approvals, expanded indications, or updated safety information. Document structures typically include fields such as basic drug information (generic name, brand name, dosage form, strength), indications, dosage and administration, contraindications, adverse reactions, drug interactions, and precautions. Units involve dosage (mg, g, IU), frequency (times/day, hours), treatment duration (days, weeks), and concentration (mg/ml, %).
Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts
High-frequency updates require the knowledge base to quickly synchronize and index information. This ensures that information referenced in multi-turn conversations remains current. Complex document structures and diverse fields necessitate more refined text segmentation strategies. This avoids truncating or obscuring critical information. For example, the drug interaction section often contains multiple complex logical relationships; its completeness must be ensured. Diverse units and dosage expressions challenge prompt parsing capabilities. The model must accurately understand the equivalence between "twice daily, 10mg each time" and "20mg daily, divided into two doses." Furthermore, the rigor of rational drug use requires the conversation system to cite data sources and update times when referencing information, enhancing credibility.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 500–800 characters | Ensures a complete knowledge point from a drug insert (e.g., indications or adverse reaction list) is contained within a single chunk. |
Recall count (Retrieval Count) | Top 5 | Covers multiple potentially relevant drugs or drug interaction information that a user query might involve. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Balances recall and precision. This avoids interference from irrelevant information while capturing subtle differences in drug names or symptom descriptions. |
maxContext | 6000–8000 tokens | Accommodates accumulated drug information, patient symptom descriptions, and medical history in multi-turn conversations, maintaining conversational coherence. |
Rerank result count (Reranked Retrieval Count) | Top 3 | Prioritizes displaying the most relevant rational drug use suggestions for the user's current question, reducing the user's screening burden. |
Knowledge Base Update Frequency | Daily | Ensures that the latest changes in drug inserts, clinical guidelines, and other sources are reflected in the knowledge base promptly. |
Three Common Mistakes
- Drug name or dosage unit recognition errors occur in conversations. This happens because prompts do not adequately consider the complexity of synonyms, abbreviations, or unit conversions.
- After multi-turn conversations, the system provides drug recommendations inconsistent with the initial question. This results from improper context management, failing to effectively reference multi-turn conversation history.
- The system cannot parse uploaded drug images, preventing it from providing rational drug use advice based on image information. This occurs because the conversation interface does not support file transfer or lacks a configured image recognition plugin.
How to Confirm Proper Configuration
- Conduct multi-turn tests for typical rational drug use scenarios (e.g., "Can drug A and drug B be taken together?", "How to medicate for symptom C?"). Check if the system's drug recommendations are accurate, complete, and consistent with the latest drug inserts.
- Randomly select updated drug information from the knowledge base. Verify if the system can cite the latest data by asking questions, and observe if the data source is correct.
- Simulate users uploading drug insert images. Verify if the system can extract key information and provide rational drug use consultations based on this information. Check the recognition accuracy of relevant fields.
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.