Data Characteristics
Contraindication and interaction data originates from drug inserts, pharmacopoeias, clinical guidelines, and specialized pharmaceutical databases. Data updates frequently, especially with new drug approvals or clinical research advancements, typically quarterly or semi-annually. Document structures are primarily structured or semi-structured, such as JSON, XML, or database records. Core fields include drug name, active ingredient, contraindications (diseases, physiological states), interacting drugs, interaction type (pharmacodynamic, pharmacokinetic), severity, and management recommendations. While dosage information often involves units like milligrams (mg) and milliliters (mL), contraindications and interactions focus more on qualitative descriptions and grading.
Constraints on Multi-turn Conversations and Prompts
The highly structured nature of contraindication and interaction data requires precise field matching during information extraction in multi-turn conversations. For example, the system must identify drug names and disease states mentioned by the user. Multi-turn conversations need to support the recognition and cross-comparison of various drug names, handling synonyms, abbreviations, and different dosage forms. Due to the criticality of interaction information, prompt design must emphasize users providing a complete and accurate list of all medications, including those currently being taken, to avoid omissions. Furthermore, the severity grading and management recommendations for interactions require prompts to clearly convey risk levels in answers and guide users to consult medical professionals, ensuring rigorous information delivery. The high frequency of data updates also requires prompts to guide the model to prioritize recalling the latest data version.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Recall count | Top 5–8 entries | Ensures coverage of potential multiple interactions or contraindications while avoiding irrelevant information interference. |
Similarity threshold | 0.75–0.85 | Improves the accuracy of recall results, filtering out irrelevant drugs or contraindication descriptions, and reducing false positives. |
Chunk size | 400–600 characters | Accommodates the common length of contraindication and interaction descriptions in drug inserts or specialized databases. |
maxContext | 6 | Retains sufficient turn history to allow users to provide multiple drug details or inquire about interactions of different drugs. |
Prompt Temperature | 0.3–0.5 | Reduces the divergence of model-generated content, ensuring accuracy and rigor in answers, and preventing fabrication of information. |
JSON Schema | Calibrate based on actual measurements | Forces the model to output structured answers, facilitating subsequent programmatic parsing and display of key information, such as interacting drugs and severity. |
Common Pitfalls
- The model fails to identify multiple drugs mentioned by the user and perform cross-queries. This occurs because the prompt does not explicitly ask the model to identify co-occurring drugs; the model only focuses on the most recently mentioned drug.
- The model omits the severity or management recommendations for interactions in its answer. This happens because the retrieved raw data's critical fields are not effectively guided by the prompt for integration and output by the model.
- Users experience long delays in receiving replies or receive garbled output after asking a question. This is because the
JSON Schemais too complex, preventing the model from generating formatted output as required, leading to parsing failures.
How to Verify Configuration
- Test different drug combinations to verify if the model accurately identifies all drugs and retrieves corresponding interaction information. Cross-reference retrieval results with the data source.
- Test drug pairs with known severe interactions. Check if the model's answer includes the severity grading of interactions and clear management recommendations. Compare these with the latest pharmacopoeia content.
- Test multi-turn conversation scenarios. Continuously ask about contraindications or interactions of different drugs. Observe if the model maintains context memory and provides coherent and accurate answers based on historical conversations.
Note: The values provided are common starting points. Measure them against your own samples for optimal performance.
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.