Data Characteristics for This Category
Dosage adjustment data primarily originates from drug inserts, clinical guidelines, pharmacopoeias, and pharmaceutical professional databases. This data typically exists as structured or semi-structured text. The update frequency is relatively low, usually coinciding with drug insert revisions or new guideline releases. Document structures are predominantly chapter-based, covering indications, contraindications, dosage and administration, and use in special populations (e.g., hepatic/renal impairment, elderly, children). Key fields include drug name, indication, patient characteristics (e.g., age, weight, liver/kidney function indicators), dosage, frequency, route of administration, maximum dose, and adjustment rationale. Units involve milligrams (mg), grams (g), milliliters (ml), times/day, and mg/kg, often with multiple units coexisting or requiring conversion.
Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts
The semi-structured nature of dosage adjustment data requires detailed entity recognition and relationship extraction during data preprocessing to effectively support subsequent Q&A. Low update frequency means that after knowledge base construction, maintenance costs are relatively manageable, but version management needs attention. Diverse patient characteristics and complex adjustment rationales dictate that during multi-turn conversations, the system must actively query for critical information to ensure recommendation accuracy. For example, when a user asks about a drug's dosage, the system needs to identify and ask about the patient's liver and kidney function, age, and other factors. The complexity of fields and units requires prompt design to explicitly instruct the model to standardize unit expressions when generating responses and to perform precise numerical calculations, avoiding vague or incorrect quantitative information. Furthermore, multi-turn conversations heavily rely on historical information, necessitating a long context window to maintain conversational coherence.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8192 or 16384 tokens | Dosage adjustment questions often involve multi-turn conversations, requiring a longer context window to accommodate patient history, medication history, and multiple follow-up questions and answers, ensuring logical coherence and accuracy. |
temperature | 0.1 – 0.3 | Dosage adjustment is a rigorous medical issue; responses need high accuracy and factuality. A low temperature helps reduce model creativity, maintaining objectivity in answers. |
top_p | 0.1 – 0.3 | Similar to temperature, top_p limits the model's selection range when generating the next word, further ensuring determinism and professionalism of the output. |
recall_threshold | 0.75 – 0.85 | This ensures that recalled knowledge snippets are highly relevant to the user's query, preventing the introduction of inaccurate or irrelevant dosage adjustment information that could affect professional judgment. |
Chunk size | 500 – 800 characters | Dosage adjustment knowledge points often contain many details. An appropriate segment length effectively preserves knowledge completeness while preventing individual segments from becoming too long and diluting core information. |
Rerank result count | 5 – 8 items | Building on recall, re-ranking further filters out the most relevant knowledge items, providing the model with richer context and improving the quality of the final response. |
Three Common Mistakes
- Dosage units are inconsistent or conversion errors occur in model responses. This happens because the prompt does not explicitly request unit standardization or the model's understanding of unit conversion is insufficient.
- The model fails to actively query critical patient physiological indicators (e.g., creatinine clearance) during the conversation, leading to incomplete dosage recommendations. This occurs because the prompt does not sufficiently emphasize the dosage adjustment decision tree or does not effectively utilize conditional judgment information from the knowledge base.
- After knowledge base migration, the model indicates the knowledge base is empty or cannot find content, showing
KnowledgeBase is emptyin the conversation results. This is usually due to a failed knowledge base index rebuild or an error during file parsing, preventing document content from being imported correctly.
How to Confirm Correct Configuration
- Test with various typical dosage adjustment scenarios. Check if the model can actively query for critical information based on patient characteristics and provide complete dosage recommendations with consistent units.
- Verify the dosage values provided in the model's response against the original data in the corresponding knowledge base documents to ensure numerical accuracy.
- Simulate medication consultations for special populations, such as patients with hepatic/renal impairment or elderly patients. Verify if the model can correctly identify and apply the appropriate dosage adjustment rules.
Note: The values provided are common starting points. Measure them against your own samples for optimal results.
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.