Data Characteristics
Pharmacovigilance data in nursing management originates from Electronic Health Records (EHR), physician order systems, nursing records, adverse event reporting systems (AEGIS), drug inserts, and clinical guidelines. This data updates frequently. Patient vital signs, medication records, and adverse reaction observations can update hourly or in real-time. Document structures are semi-structured and unstructured, including free-text nursing observations, doctor diagnoses, and medication orders. Key fields include drug generic name, batch number, dosage, administration route, administration time, patient ID, allergy history, adverse reaction symptom description, onset time, severity assessment (e.g., CTCAE grade), intervention measures, and outcomes. Units include mg, ml, μg/kg/min, and times/day.
Constraints Imposed by These Characteristics on Multi-turn Conversation and Prompts
High data update frequency requires the knowledge base to support rapid synchronization and indexing. This ensures the conversation model accesses the latest information. Semi-structured and unstructured text features necessitate strong text understanding and entity extraction capabilities. This accurately identifies critical information like drugs, symptoms, and event times. In multi-turn conversations, patient medication history and allergy history require continuous maintenance. This supports precise pharmacovigilance judgments. Ambiguous symptom descriptions require prompt designs that guide the model to ask follow-up questions from multiple angles or match symptoms against the knowledge base. Dosage precision constrains the conversation model. Medication recommendations must strictly adhere to numerical values and units, avoiding misjudgments.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 6 turns | Retains recent medication and adverse reaction observation context, balancing efficiency and accuracy. |
Chunk size (Segment Length) | 800 characters | Accommodates detailed symptom descriptions and observation logs in nursing records, ensuring semantic completeness. |
Recall count (Recall Count) | 10 items | Covers relevant drug inserts, adverse reaction cases, and patient history, increasing recall comprehensiveness. |
Similarity threshold (Similarity Threshold) | 0.75 | Filters out irrelevant knowledge chunks, improving the precision of recall results. |
Rerank result count (Reranked Return Count) | 5 items | Selects the most relevant knowledge snippets for the current conversation, reducing the model's burden of processing irrelevant information. |
MODEL_MAX_TOKENS | 4096 | Accommodates complex medication regimens and multiple adverse reaction descriptions, ensuring model processing capability. |
Common Pitfalls
- The conversation model fails to cite knowledge base content in its responses. This results in generic or inaccurate replies. The cause is insufficient knowledge base recall effectiveness or prompts that do not explicitly instruct the model to cite.
- During multi-turn conversations, the model forgets patient historical medication or allergy history. This leads to inappropriate recommendations. The
maxContextsetting is too small, failing to retain sufficient conversation history. - The model generates medically incorrect numerical values or units when handling user questions about drug dosage or administration routes. This manifests as incorrect numbers or unit confusion. The cause might be incorrect parsing of relevant fields in the knowledge base or prompts that do not emphasize strict matching of numerical values and units.
Verification Steps
- Conduct simulated conversation tests. Ask about adverse reactions and management for specific drugs. Check if the model's response accurately cites information from the knowledge base's drug inserts.
- Simulate a patient mentioning different medication situations multiple times. Observe if the model correctly identifies and associates previous medication information in subsequent conversations, such as asking about allergy history.
- Input queries containing ambiguous symptom descriptions. See if the model guides the user to provide more specific information through multi-turn conversations and ultimately provides reasonable suggestions based on the knowledge base.
Note: The values provided are common starting points. Measure them against your 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.