Data Characteristics
Quality documentation for home medical devices originates from product design and development, manufacturing, risk management, clinical evaluation, and post-market surveillance. These documents have a relatively stable update frequency, typically aligning with product lifecycles, regulatory revisions, and quality system audit cycles (e.g., annual audits or major changes).
Document structures are rigorous, often presented in standardized templates or formats. Examples include ISO 13485 Quality Management System documents, Technical Documentation (TD), Design History Files (DHF), Risk Management Files (RMF), user manuals, instructions for use, and batch production records. Fields and units are highly specialized, containing medical device-specific terminology, measurement units (e.g., millimeters, volts, milliamperes, pascals), standard codes (e.g., UDI, GMDN codes), and specific test parameters and results.
Constraints on Knowledge Base Retrieval and Recall
The standardized structure and specialized terminology of home medical quality documents require precise context preservation during knowledge base document chunking. This avoids misinterpretation and distorted retrieval results.
Stable update frequencies allow for periodic full or incremental update strategies during index building, ensuring knowledge timeliness. Unique measurement units and standard codes challenge the accuracy of embedding generation by vector models. Models must understand the semantic associations of these specialized symbols, preventing simple character matching failures.
The prevalence of multi-level directories and nested documents requires the knowledge base to parse complex document structures. This ensures all hierarchical content is effectively retrievable.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 500–800 characters | Ensures individual chunks contain sufficient context while avoiding excessive length that could lead to information redundancy and reduced recall precision. |
Chunk Overlap Length (Chunk Overlap) | 50 characters | Maintains semantic continuity between chunks, especially for specialized terminology and process descriptions. |
Recall count (Recall Count) | top 5 | Balances retrieval efficiency with comprehensive recall, covering key information and reducing interference from irrelevant results. |
Similarity threshold (Similarity Threshold) | Calibrate based on actual measurements | Adjust based on actual retrieval performance and document characteristics to avoid missed or false recalls. |
Rerank result count (Rerank Return Count) | top 3 | Further enhances the relevance of the most pertinent document segments using a reranking model after initial recall. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Addresses the parsing requirements of large technical files and complex PDF documents, preventing timeouts. |
Common Pitfalls
- The knowledge base fails to correctly parse multi-level directories and image content within Feishu documents. This results in critical information not being retrievable. The parser's limited ability to handle non-standard formats or embedded objects causes this.
- The system behavior deviates from expectations when users selectively skip knowledge base retrieval. This may be due to incorrect configuration of control variables like
skipKnowledgeBaseor the backend service not recognizing them. - Switching retrieval sources based on the input knowledge base name results in generic error messages like
{ "message": "common:core.chat" }. This typically happens when the dynamically switched knowledge base ID does not map to an actual knowledge base instance, or permission configurations are incorrect.
Verification Steps
- Upload a typical home medical device technical file with multi-level directories, images, and specialized terminology to the knowledge base. Perform a retrieval to check if all levels and objects are correctly recalled.
- Retrieve using a document containing specific measurement units and standard codes. Check if these specialized details are accurate in the recall results. Observe result changes by adjusting the
Similarity threshold(Similarity Threshold). - Execute a series of retrieval requests with parameters, such as explicitly specifying or skipping a knowledge base. Verify if system behavior aligns with the expected outcomes of parameters like
skipKnowledgeBaseorkbId. - Perform small-scale incremental updates to the knowledge base regularly. Observe if retrieval results immediately reflect the latest content, confirming the update mechanism is effective.
The values provided are common starting points. Measure them against your own samples for optimal results.
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.