Data Characteristics
Telemedicine product data primarily comes from manuals, operation guides, technical white papers, compliance certification documents, and clinical application guidelines for medical devices, diagnostic reagents, and remote monitoring equipment. These documents are typically in PDF format, with a small number in Word or image formats. Update frequency for product manuals and technical specifications is periodic, usually quarterly or semi-annually, driven by product iterations, regulatory updates, or new feature releases. Document structures are clearly hierarchical, containing numerous charts, specialized terminology, dosage units, operating procedures, and safety warnings. Fields include product models, batch numbers, expiration dates, storage conditions, usage instructions, contraindications, and adverse reactions. They strictly adhere to medical industry standards; for example, measurement units must be precise to multiple decimal places.
Constraints Imposed by These Characteristics on Document Parsing and Chunking
The clear hierarchy and dense specialized terminology in telemedicine product documents require the parser to accurately identify chapter titles, subtitles, and logical relationships between paragraphs to avoid information silos. The abundance of charts and images necessitates OCR support or descriptive extraction of image content to ensure information completeness. Frequent updates mean the knowledge base must support incremental updates and version management, ensuring retrieved information is always current and compliant. Strict measurement units and specialized fields require that critical values or terms are not arbitrarily truncated during chunking; for example, "2.5mg/mL" must be treated as a single entity. Additionally, key information such as safety warnings and contraindications needs higher weighting or independent, complete retention within chunks to prevent potential medical risks.
Configuration Recommendations
| Configuration Item | Recommended Value | Rationale for this Value |
|---|---|---|
Chunk size (Chunk Length) | 800–1200 characters (characters) | Retains sufficient contextual information while preventing individual chunks from becoming too large, which can degrade retrieval efficiency. |
Chunk Overlap Length (Chunk Overlap Length) | 100–200 characters (characters) | Ensures semantic continuity between adjacent chunks, especially important for specialized terminology and operational procedure descriptions. |
Chunking Rule | Split by Title, Paragraph | Prioritizes identifying the logical structure of the document to maintain semantic integrity and avoid truncating critical information. |
Skip Headers/Footers | Enabled (Enable) | Excludes irrelevant information that could interfere with semantic understanding, improving retrieval accuracy. |
Enable Image OCR | Enabled (Enable) | Extracts key information from charts, such as product specifications and dosage diagrams, enriching the knowledge base content. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Addresses the parsing time requirements for complex documents (e.g., those with many charts or multi-level directories). |
Common Pitfalls
- Document parsing timeouts, appearing as files remaining in a processing state for an extended period after upload or directly reporting
PARSE_FILE_TIMEOUTerrors. This usually occurs because thePARSE_FILE_TIMEOUT_SECONDSconfiguration is set too low to accommodate the parsing time for large or complex documents. - Critical numbers or specialized terms are truncated in recall results after knowledge base chunking, such as incomplete dosage units. This happens when the
Chunk size(Chunk Length) is set too small, preventing the model from semantically recognizing and retaining complete entities. - After upgrading the FastGPT version, some CSV files fail to chunk correctly, reporting
Cannot redefine property: toString. This may be due to optimizations in the new version's CSV parsing logic, where older file formats or content contain incompatible special characters or encoding issues.
How to Verify Correct Configuration
- Upload multiple typical documents (e.g., product manuals, technical white papers) and check if the chunking results completely retain key information, such as product models, dosages, and contraindications.
- Randomly select chunks and check if there is sufficient overlap content between adjacent chunks to ensure contextual coherence.
- For documents containing charts, check if image OCR results correctly identify and extract text information from the images, especially critical text in data tables or flowcharts.
- Simulate user queries to retrieve specific product usage methods or adverse reactions, evaluating the relevance and completeness of the recalled chunks, ensuring critical safety information can be accurately retrieved.
The values provided are common starting points. They 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.