Knowledge Base Retrieval and Recall for Ophthalmic

Data Characteristics

Core data sources for ophthalmic registration and declaration documents include clinical trial reports, non-clinical drug study reports, product inserts, registration certificates, various guidelines, and international/domestic pharmacopoeia standards. These documents have a relatively low update frequency, typically changing with the drug's lifecycle or regulatory revisions, such as the release of new pharmacopoeias or updated guidelines. Document structures are highly standardized, adhering to CTD (Common Technical Document) formats from regulatory bodies like China's National Medical Products Administration (NMPA), the U.S. Food and Drug Administration (FDA), or the European Medicines Agency (EMA). They contain detailed chapters and sub-sections. Data fields cover drug active ingredients, indications, dosage and administration, adverse reactions, pharmacokinetics, and pharmacodynamics. Units strictly follow the International System of Units (e.g., milligrams, milliliters, micromoles) or specific medical measurement units (e.g., diopters, mmHg for intraocular pressure).

Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall

The standardized document structure and low update frequency of ophthalmic registration and declaration documents require the knowledge base to effectively parse CTD formats during construction, ensuring semantic integrity at the chapter level. The strictness of fields and units dictates that retrieval results must be precise, with low tolerance for error. For example, when querying "intraocular pressure," the system should differentiate between "mmHg" and "kPa" and perform unit conversion or provide a prompt. The volume of such data is often large, with a single document potentially spanning thousands of pages, posing challenges for the knowledge base's chunking strategy and indexing efficiency. Additionally, due to numerous specialized terms and strong context dependency, simple keyword matching can lead to inaccurate recall. For instance, searching for "glaucoma" may require linking to its treatment drugs, clinical manifestations, and relevant diagnostic standards. This necessitates the knowledge base possessing deeper semantic understanding capabilities to precisely identify and recall relevant snippets from vast amounts of text.

Configuration Recommendations

Configuration Item	Suggested Value	Rationale
`Chunk size` (Chunk Length)	`800–1200 characters`	Registration and declaration documents have long paragraphs; maintaining contextual integrity avoids semantic fragmentation.
`Recall count` (Number of Retrieved Chunks)	`Top 8–12 chunks`	Ensures coverage of multiple potentially relevant paragraphs, balancing efficiency and recall breadth.
`Similarity threshold` (Similarity Threshold)	Calibrated by actual measurement	Based on specific datasets and query types, balances recall and precision. An initial value of `0.75` can be used.
`Rerank result count` (Number of Reranked Chunks)	`Top 5 chunks`	Optimized by a reranking model to focus on the most relevant results, reducing user reading burden.
`chunkOverlapRatio`	`0.1`	A small overlap helps connect semantics across chunks, preventing information loss.
`embeddingModel`	`text-embedding-ada-002`	Suitable for specialized domain texts, with strong semantic understanding capabilities.

Three Common Pitfalls

Too few or no retrieval results: This occurs when the Similarity threshold (Similarity Threshold) is set too high, strictly filtering out potentially relevant but slightly less similar document snippets.
Incomplete recalled paragraphs or lack of context: This happens when the Chunk size (Chunk Length) is set too short, splitting closely related text and disrupting semantic coherence.
The system returns a large number of irrelevant results: This phenomenon occurs when Recall count (Number of Retrieved Chunks) is too high and Rerank result count (Number of Reranked Chunks) is not effectively configured, leading to the presentation of many low-relevance items.

How to Confirm Proper Configuration

Select a batch of representative ophthalmic registration and declaration queries. Check if the recall results contain key information points and verify the distribution of Similarity Scores.
Validate the system's retrieval for specific terminology (e.g., "retinopathy," "phacoemulsification"), ensuring it accurately returns relevant definitions, clinical guidelines, or experimental data.
Simulate user queries and observe whether the returned document snippets can independently answer the question or provide sufficient context for further reading, evaluating the reasonableness of Chunk size (Chunk Length).

The values provided are common starting points and should be measured against specific 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.