Data Characteristics
Literature support data for Medical Information (MI) responses comes from global biomedical journals, clinical guidelines, conference abstracts, drug inserts, and internal medical databases. Update frequencies vary: journal articles and conference abstracts might update monthly or even weekly, while drug inserts and clinical guidelines are more stable, typically revised quarterly or annually. Literature data mixes structured and unstructured forms, usually including standard fields like title, author, abstract, introduction, methods, results, discussion, and references. Document lengths vary significantly, from short conference abstracts of a few hundred characters to review articles tens of thousands of characters long. Content often contains complex medical terminology, drug names, disease codes (e.g., ICD-10), gene sequences, compound structures, and dosage units (e.g., mg/kg, IU).
Constraints on Vector Models and Indexing
The diversity and complexity of literature data place specific demands on vector models and indexing strategies. First, highly specialized medical terminology requires vector models to have strong semantic understanding. They must accurately capture the specific meaning of terms in the biomedical field, differentiate synonyms and near-synonyms, and handle abbreviations. Second, significant variations in document length necessitate flexible chunking strategies. These strategies must avoid excessively long chunks that dilute key information and overly short chunks that lose contextual relevance. Effective indexing of long documents requires considering hierarchical structures, such as independent vectorization of chapters, paragraphs, or even figure captions. Furthermore, inconsistent data update frequencies require the indexing system to support incremental updates and version management, ensuring retrieved information is always current. Finally, the need for traceability—identifying the specific paragraph in a document that supports an answer—means index design must preserve original document block-level metadata.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
embeddingModel | DeepSeek-v2 or bge-large-zh-v1.5 | These models perform well in semantic understanding and specialized terminology processing for the Chinese biomedical domain. |
Chunk size (Chunk Length) | 500–800 characters (characters) | Balances contextual completeness with retrieval efficiency, avoiding dilution of key information by overly long chunks or loss of semantics by overly short ones. |
Chunk overlap (Chunk Overlap) | 50–100 characters (characters) | Ensures information at chunk boundaries is not fragmented, improving retrieval recall. |
Recall count (Recall Count) | 8–15 entries (items) | Considering that medical questions often involve multiple aspects, increasing the recall count helps cover more comprehensive information. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Ensures the relevance of retrieval results while avoiding omission of potentially critical literature segments. Calibrate based on actual measurements. |
Rerank result count (Reranked Return Count) | 3–5 entries (items) | After reranking, focus on the most relevant few items to improve the precision of the final answer. |
Common Pitfalls
- Retrieval results contain many irrelevant or low-relevance literature segments. This usually happens when the
Similarity threshold(similarity threshold) is set too low, leading to the recall of semantically loosely related document blocks. - Answer content cannot be accurately traced back to specific literature sources. This indicates that during knowledge base construction, the original document's
metadataorchunk_idwas not correctly associated or stored. - When processing long documents, the model provides incomplete answers or misses key information. The reason might be that the
Chunk size(chunk length) is set too short, causing important context to be fragmented and not fully vectorized.
How to Confirm Proper Configuration
- Select several typical biomedical literature pieces as a test set. Perform question-answering tests and check if the answers include key information from these documents.
- Verify whether the literature traceability links or identifiers provided in the answers accurately point to specific chapters or paragraphs of the original documents.
- Adjust the
Similarity threshold(similarity threshold) parameter. Observe changes in the relevance of recalled literature segments until recall rate is maintained while reducing interference from irrelevant information. - Monitor the knowledge base's incremental update mechanism. Ensure newly published medical literature is timely indexed and included in retrieval, and updates to older literature versions are correctly reflected.
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.