Ophthalmic Pharmacovigilance: Multi-turn Conversations and

Ophthalmic Data Characteristics

Ophthalmic pharmacovigilance data originates from clinical trial reports, real-world evidence (RWE), post-market surveillance systems (e.g., FDA FAERS, EMA EudraVigilance), and electronic medical records from healthcare institutions. Data update frequencies vary. Clinical trial data typically releases centrally after study completion, while post-market surveillance data flows in continuously and fragmentedly. Document types are diverse, including structured reports (e.g., CIOMS I forms), semi-structured medical records (e.g., handwritten doctor's notes, patient interview records), and unstructured scientific literature. Fields include patient demographics, medication history, adverse event descriptions (including signs, symptoms, diagnosis), event onset time, outcome, and causality assessment. Adverse event descriptions often involve ophthalmic-specific terminology and anatomical sites, such as "retinal detachment," "macular edema," "elevated intraocular pressure." Units involve vision (e.g., Snellen chart) and intraocular pressure (mmHg).

Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts

The diversity of ophthalmic pharmacovigilance data sources and varying update frequencies require the knowledge base to continuously integrate new information and maintain awareness of the latest data during multi-turn conversations. The prevalence of semi-structured and unstructured documents makes precise information extraction and entity recognition challenging, especially when dealing with complex ophthalmic terminology and symptom descriptions. In multi-turn conversations, users may gradually supplement or correct adverse event details. The system must accurately understand the context and infer information based on historical dialogue. For example, a user might initially mention "blurred vision" and later add "accompanied by floaters, appeared after medication." The system needs to link this information to specific drugs and adverse reactions. The presence of ophthalmic-specific fields and units requires prompt design to fully consider the semantic understanding of these professional terms and the processing of quantitative information, avoiding misjudgments due to ambiguous terminology or confusing units.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`800–1200 tokens`	Accommodates the detailed nature of ophthalmic adverse reaction descriptions and balances the model's ability to process complex contexts, avoiding redundant information interference.
`temperature`	`0.3–0.5`	Ensures the stability and accuracy of generated responses, reduces hallucinations, and meets the strict requirements of pharmacovigilance.
`top_p`	`0.7–0.8`	Further constrains the model's output range, focusing on highly relevant information and reducing the generation of irrelevant information.
`System Prompt`	See below	Presets roles and guidance, ensuring the model understands the professional context and task objectives of ophthalmic pharmacovigilance.
`Knowledge base recall count`	`Top 5–8 entries`	Balances recall breadth and precision, ensuring coverage of multi-faceted information while avoiding the introduction of excessive noise.
`Similarity threshold`	`0.75–0.85`	Accurately matches professional terminology and symptom descriptions in ophthalmic adverse reaction reports, filtering out irrelevant knowledge snippets.

Common Pitfalls

Dialogue includes prompts like "Unable to find relevant drug adverse reaction information" or "Reply content does not match ophthalmic symptoms." This indicates a lack of specific ophthalmic drugs or rare adverse reactions in the knowledge base, or that document segmentation is too coarse, diluting critical information.
After multi-turn conversations, the model misunderstands the patient's ophthalmic symptom description, for example, confusing "dry eyes" with "vision loss." This occurs because the prompt does not explicitly guide the model to differentiate similar symptoms, or the model's context window is too small, leading to the loss of early key information.
When a user asks about contraindications for a specific ophthalmic drug, the model provides a generic response, failing to offer targeted ophthalmic contraindications. This happens because drug instruction documents in the knowledge base are not effectively parsed, or the prompt fails to guide the model to extract structured contraindication information.

Validation Steps

Select 10–15 typical ophthalmic adverse reaction reporting scenarios and conduct multi-turn conversation tests. Check if the model can accurately extract key information and provide compliant responses. Compare test results with expected answers to evaluate accuracy.
Input queries containing ophthalmic professional terminology (e.g., "vitreous opacity," "fundus hemorrhage") and quantitative units (e.g., "intraocular pressure 25 mmHg"). Check if the model can correctly understand and cite knowledge base document snippets containing this information, evaluating recall quality.
Simulate scenarios where users gradually supplement adverse event details. Observe the model's understanding and reasoning abilities across different turns, ensuring accurate information integration. Use human evaluation to determine coherence and logical consistency.
Check maxContext usage after each conversation to ensure that in complex dialogue scenarios, the context window can accommodate enough information without performance degradation due to excessive length. Adjust the maxContext value based on actual conditions.

The values provided are common starting points and should be measured against the reader's 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.