Data Characteristics
Home medical device pharmacovigilance data originates from user-initiated reports, manufacturer after-sales records, and public regulatory reports. Data update frequencies vary: user reports can be real-time, manufacturer records are often batched or event-summarized, and regulatory reports have fixed publication cycles (e.g., quarterly or annually).
Document structures differ. User reports are mostly unstructured free-text descriptions, including symptoms, device models, and usage environments. Manufacturer records are often semi-structured, involving problem categories, processing workflows, device serial numbers, and batch numbers. Regulatory reports are more standardized, typically containing event IDs, device names, adverse event types, occurrence dates, and outcome results.
Common units in the data include time units (days, hours), quantity units (times, units), and descriptive terms. Complex measurement units are rare.
Constraints on Deployment and Upgrade
The highly unstructured and semi-structured nature of home medical device pharmacovigilance data demands advanced text parsing and entity recognition capabilities from FastGPT during deployment. Non-real-time data sources require knowledge base construction and update strategies to handle both batch imports and incremental updates, avoiding frequent full rebuilds.
Colloquialisms and typos in free-text descriptions challenge the robustness of tokenizers and fuzzy matching. Key identifiers like device models and batch numbers require careful index granularity during knowledge base index construction to support precise retrieval.
Local or specific environment deployments need to support parsing various data formats (e.g., CSV, PDF, Word documents) and comply with anonymization requirements for sensitive information. During upgrades, key considerations include new model compatibility, smooth migration of old knowledge bases, and efficient historical data index reconstruction.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 3000 Tokens | Balances completeness of user descriptions with model processing efficiency, preventing context overflow. |
Chunk size (Segment Length) | 500 characters | Adapts to unstructured text characteristics, ensuring completeness of single segments. |
Recall count (Recall Count) | 8 entries | Increases relevant information coverage, addressing ambiguity in colloquial descriptions. |
Similarity threshold (Similarity Threshold) | 0.75 | Appropriately relaxes matching strictness while ensuring recall accuracy. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Processing large PDF or Word format regulatory reports can be time-consuming. |
UPLOAD_FILE_MAX_SIZE | 500 MB | Allows uploading documents containing numerous charts or reports. |
Common Mistakes
- Model answers still cite old data after a knowledge base update: The knowledge base index was not fully rebuilt, or the cache was not refreshed in time.
- User-submitted device model identification fails, preventing association with relevant knowledge: Text parsing configuration is not optimized for specific naming rules, or the entity recognition dictionary is outdated.
- Thinking process displays abnormally or cannot be collapsed after deploying a local model: The
THINKING_PROCESS_ENABLEDparameter configuration does not match front-end rendering logic, or the local model API response format is inconsistent.
Verification Steps
- Upload adverse event reports for home medical devices in various formats. Check if knowledge base document parsing is normal and if segment content is reasonable.
- Simulate user questions including common device models, symptom descriptions, and usage scenarios. Verify if model answers accurately recall relevant knowledge.
- After updating the knowledge base, perform an incremental or full index rebuild. Verify if new data can be effectively retrieved and referenced.
The values provided are common starting points. Measure them against your own samples to determine optimal settings.
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.