Data Characteristics
Rehabilitation equipment, as a critical component of medical devices, requires highly standardized and professional quality documentation. Data originates from regulations and standards published by the National Medical Products Administration (NMPA), technical guidelines from industry associations, internal product design documents, production process specifications, inspection records, clinical evaluation reports, and post-market surveillance data. These documents are typically in PDF, Word, or Excel formats. Content includes technical parameters, performance indicators, risk management, clinical trial data, user manuals, and maintenance instructions. Document updates are irregular but critical, driven by regulatory revisions, new product launches, technological iterations, and quality incidents. Document structures are rigorous, often including chapter numbers, appendices, figures, and tables. Fields include Unique Device Identification (UDI), registration certificate numbers, product models, serial numbers, and production batches. Units strictly adhere to the International System of Units (SI) and medical device industry conventions.
Constraints from These Characteristics on Citation and Traceability
The standardization and professionalism of rehabilitation equipment quality documents demand precise and traceable citations. Irregular regulatory updates require the knowledge base to support efficient incremental updates and version management, ensuring cited content is current and valid. The prevalence of figures and tables in documents challenges text extraction and semantic understanding; traditional pure-text RAG methods may struggle with accurate parsing. Unique fields like UDI and registration certificate numbers necessitate precise matching during citation to avoid confusion between different product models or batches. Unit strictness requires avoiding unit conversion errors during information extraction to ensure data consistency. These constraints collectively point to a core requirement: cited content must be semantically relevant and possess high authority and verifiability to meet inspection and compliance requirements.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk Size | 500–800 characters | Balances semantic completeness and recall efficiency, avoiding redundant information in long paragraphs. |
Overlap Size | 50–100 characters | Ensures contextual continuity and addresses cross-paragraph semantic dependencies. |
Recall Count | Top 8–12 | Covers potentially relevant information, balancing recall precision and processing overhead. |
Similarity Threshold | 0.75–0.85 | Ensures recalled content is highly relevant to the query, reducing low-quality citations. |
Rerank Count | Top 3–5 | Selects the most relevant content, improving the accuracy of the final answer. |
File Type Whitelist | pdf, docx | Focuses on mainstream quality document formats, improving parsing success rates. |
Common Pitfalls
- Symptom: The model's response cites outdated or irrelevant regulations or technical standards. Reason: Knowledge base documents are not updated promptly, or chunking strategies lead to incorrect recall of old version information.
- Symptom: The model cites knowledge base content, but the answer lacks critical product model or registration certificate number information. Reason: Structured fields are not effectively identified during text extraction, or citation granularity is too broad, diluting key information.
- Symptom: After uploading many quality documents, knowledge base construction or query response speed significantly slows down. Reason: The
PARSE_FILE_TIMEOUT_SECONDSparameter is set too low, or file parsing concurrency is insufficient, leading to processing bottlenecks.
Verification Steps
- Select representative queries. Check if the source documents cited in the model's response are the latest versions and verify their accuracy.
- For queries containing product models, UDI, or registration certificate numbers, verify that these key fields in the model's response exactly match the cited sources.
- Upload documents with complex figures and tables. Test if the model can accurately extract and cite data from tables, such as equipment parameters or test results.
- Simulate an inspection scenario. Use compliance-related queries to assess if the model can cite relevant regulatory clauses and internal control documents, and verify the traceability of citations.
The values provided are common starting points and should be measured against your 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.