Ophthalmic Pharmacovigilance: Citation and Traceability

Data Characteristics

Ophthalmic pharmacovigilance data originates from clinical trial reports, real-world evidence (RWE) studies, post-market surveillance reports, medical literature, patient reporting systems, and regulatory adverse drug reaction (ADR) databases. This data updates frequently, especially when new drugs launch or new serious adverse reaction signals emerge. Document structures vary, including unstructured text reports (e.g., patient narratives, physician notes), semi-structured case report forms (CRFs), and structured database entries.

Fields include general information such as drug name, batch number, dosage, administration route, adverse reaction name, onset time, severity, and outcome. Additionally, eye-specific signs (e.g., vision loss, elevated intraocular pressure, conjunctival hyperemia), ophthalmic examination results (e.g., slit lamp examination, fundus examination, visual field examination), and concomitant ophthalmic medications are present. Units involve vision (Snellen chart, logMAR chart), intraocular pressure (mmHg), and visual field (degrees), often accompanied by qualitative descriptions.

Constraints on Citation and Traceability

The multi-source and heterogeneous nature of ophthalmic pharmacovigilance data challenges unified citation management and traceability. Extracting key information from unstructured text reports is difficult, potentially leading to fragmented or missing citations. Semi-structured and structured data are easier to extract, but their specialized and diverse fields require the RAG system to accurately identify and link them.

High update frequency necessitates frequent knowledge base synchronization to ensure citation timeliness and accuracy. The presence of eye-specific fields requires precise matching of relevant medical terms during retrieval and citation, preventing incorrect citation of non-ophthalmic adverse reaction information. Furthermore, adverse reaction reports often contain sensitive patient information, requiring data anonymization and compliance with privacy regulations during citation. Traceability must link back to specific sections, page numbers, or database record IDs of original reports to support subsequent medical review and validation.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`Chunk size`	`500–800 characters`	Balances the detailed descriptions in ophthalmic case reports with the RAG model's processing capacity, preventing excessive truncation that leads to context loss.
`Chunk Overlap Length`	`100–150 characters`	Ensures contextual continuity between paragraphs, especially when describing the progression of adverse reactions.
`Recall count`	`Top 8 entries`	Considers the complexity of ophthalmic adverse reactions and potential multiple related factors, increasing recall to improve coverage of relevant information.
`Similarity threshold`	`0.75–0.85`	While ensuring recall accuracy, this range appropriately broadens the threshold to capture synonyms or related concepts for medical terms, particularly ophthalmic professional vocabulary.
`Rerank result count`	`Top 3 entries`	After re-ranking model optimization, focuses on the most relevant and information-dense citation snippets, enhancing the precision of the final response.
`Maximum citationtokenNumber`	`1500 token`	Limits the overall length of cited text, preventing overly long citations from diluting effective information, while conforming to the context window limits of mainstream large models.

Common Mistakes

The knowledge base retrieves relevant adverse reaction reports, but the AI response only provides a conclusion without specific citation sources. This typically occurs because the Rerank result count in the RAG configuration is too low, or the Maximum citationtokenNumber limit is too strict, causing effective citation snippets to be filtered before final presentation.
When retrieving ophthalmic adverse reactions, the AI cites a large amount of general adverse reaction information unrelated to the eyes. This may be because the Similarity threshold is set too low, or the knowledge base preprocessing did not effectively distinguish between eye-specific and general medical terms.
A user asks about a rare adverse reaction to a specific ophthalmic drug, and the system fails to recall relevant information, even if the knowledge base contains it. This may be related to the Chunk size being set too short, causing the complete description of the rare adverse reaction to be cut into discontinuous fragments, affecting retrieval matching.

How to Confirm Correct Configuration

For typical ophthalmic adverse drug reaction queries, check whether the report ID, page number, or paragraph cited in the AI response accurately points to the original source in the knowledge base.
Enter queries containing ophthalmic professional terms (e.g., "hypopyon," "vitreous opacity") and verify that the recalled citation snippets precisely include these terms and their related descriptions.
Select several lengthy ophthalmic clinical trial reports and simulate questions about key adverse reaction events within them. Check whether the AI's cited content covers the complete event description without excessive truncation.

The values given 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.