Data Characteristics
Pharmacovigilance quality documents include Adverse Drug Reaction (ADR) reports, safety updates, risk management plans, post-marketing safety study protocols, and Pharmacovigilance System Master Files (PSMF). These documents originate from various sources: clinical trial data, real-world data, regulatory submissions, literature searches, and internal safety databases. Documents update frequently, especially after new drug launches or when new safety information emerges. Document structures are complex, often containing extensive unstructured text like adverse event descriptions, medical terminology, and diagnostic results. They also contain structured fields such as drug names, dosages, event times, reporter information, and patient demographics. Units involved include dosage units (mg, g, ml) and time units (days, months, years).
Constraints on Deployment and Upgrades
Pharmacovigilance document characteristics impose specific requirements on FastGPT deployment and upgrades. High update frequency necessitates efficient incremental update mechanisms for the knowledge base, ensuring timely retrieval of the latest safety information. Complex document structures require FastGPT to have robust multi-format file parsing capabilities during data ingestion, especially for accurate extraction from unstructured text like PDFs and Word documents. Accurate identification of medical terminology and specialized fields is critical, as it directly impacts recall precision. During upgrades, due to data sensitivity and frequent updates, data migration must maintain completeness and consistency, preventing information loss or version conflicts. Deployment environments also need sufficient computing resources to handle large-scale text vectorization, retrieval requests, and high concurrent access.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 500 MB | Accommodates large PSMF files or bulk report uploads |
maxContext | 6000 token | Covers the complete context of typical pharmacovigilance reports |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Accounts for parsing time of complex PDF and Word documents |
Chunk size | 800–1200 characters | Balances semantic completeness with recall granularity for medical text |
Recall count | Top 10 entries | Increases coverage of relevant information for complex queries |
Similarity threshold | Calibrate by measurement, e.g., 0.75 | Ensures precision in specialized terminology matching, avoids irrelevant information |
Common Pitfalls
- Key safety information becomes unretrievable after knowledge base updates, resulting in empty or inaccurate query results. This occurs due to file parsing failures or incorrect incremental update strategy configuration, preventing new document content from being properly ingested or indexes from being rebuilt promptly.
- After a system upgrade, the recognition of specialized vocabulary significantly degrades, leading to lower quality answers. This may be due to changes in model compatibility with domain-specific terms in the new version, or improper migration of custom dictionaries during the upgrade.
- System response is slow or times out when processing large batches of adverse event reports. This typically indicates insufficient resource allocation, such as
CPU,RAM, orIOPS, failing to meet performance requirements for high-concurrency vectorization and retrieval.
Verification Steps
- Upload a test document containing the latest adverse event information. Then, perform relevant queries and verify if the query results include the critical safety information from that document.
- Use a set of test questions containing specialized medical terms and drug names. Verify that the retrieved document segments are accurate and cover the core points of the questions, and assess if semantic matching meets expectations.
- Simulate high-concurrency user access scenarios. Observe system response times when handling multiple query requests, ensuring stable operation within defined performance thresholds.
The values provided are common starting points and 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.