Data Characteristics
Quality documents in telemedicine primarily originate from regulatory files issued by health commissions, hospital-internal Standard Operating Procedures (SOPs), risk management manuals, training materials, and patient feedback records. These documents update frequently, especially policy and regulation files, which often undergo multiple revisions or additions annually. Document structures are typically PDF regulations, Word implementation details, or structured JSON data. Content covers medical quality management systems, diagnostic and treatment guidelines, equipment usage instructions, and data security and privacy protection clauses. Fields and units are highly specialized, such as "diagnosis code ICD-10," "drug dosage mg/kg," and "service duration minutes," demanding extreme accuracy.
Constraints on Knowledge Base Retrieval and Recall
Frequent updates to telemedicine quality documents require the knowledge base to have efficient incremental update and version management capabilities to ensure the timeliness of retrieval results. The diversity of document formats, particularly complex charts and tables in PDFs and Word files, poses challenges for text extraction and segmentation, potentially leading to semantic misunderstandings. Highly specialized fields and units require the embedding model to accurately capture their medical meaning during vectorization, preventing incorrect recalls due to literal similarities. Furthermore, document content often involves multi-party responsibility allocation and operational procedures, demanding high contextual coherence and completeness. Single short text segments can fragment important information, affecting final retrieval quality.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 800–1200 characters | Telemedicine documents often contain detailed operational steps and descriptions. Longer chunks help maintain contextual integrity and reduce the risk of critical information being cut off. |
Chunk Overlap Length (Chunk Overlap Length) | 100–200 characters | Appropriate overlap length helps maintain semantic coherence at chunk boundaries, especially when describing processes or complex concepts. |
Recall count (Recall Count) | Top 5 | Queries for quality documents typically require precise matching. Recalling too many items can introduce noise and increase subsequent processing burden. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | The specialized nature of telemedicine documents requires a higher similarity threshold to ensure the relevance of retrieval results and avoid interference from low-relevance documents. |
Rerank result count (Rerank Return Count) | 3 | Building on high-quality recall, reranking further optimizes sorting, prioritizing the most relevant core content. |
MAX_FILE_SIZE_MB | 100 MB | Considering PDF documents containing numerous charts and tables, a larger file upload limit supports the import of complete documents. |
Common Pitfalls
- After a knowledge base update, retrieval results do not reflect the latest policies or procedures. This occurs because the knowledge base's re-embedding or index update was not triggered in time, leading to retrieval based on outdated data.
- When a user queries for drug usage and dosage, the retrieval results contain many irrelevant disease introductions. This may be due to overly short chunks or an embedding model that fails to distinguish between medical entities and general concepts, resulting in imprecise vectorized semantics.
- After importing an Excel file with multiple tables, retrieval results are chaotic or critical data is missing. This manifests as a system error
undefined model must match "^(text"or incorrect chunking. The cause is a lack of preprocessing for tabular data or an inappropriate automatic chunking strategy, leading to the loss of table structural information.
Validation
- Upload the latest policy documents. Retrieve using keywords and sentences. Verify that retrieval results include new policy content and check the cited document version numbers.
- For queries involving specialized terminology and diagnostic codes, check if retrieval results accurately point to relevant document sections. Verify that fields and units within are correctly matched.
- Ask questions about documents containing complex tables. Confirm that retrieval results correctly identify key information within the tables, such as drug dosages and examination items, and verify their accuracy.
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.