Data Characteristics
Hematologic oncology pharmacovigilance data typically includes information from clinical trial reports, real-world observational studies, medical literature, adverse event reporting systems (such as FAERS, EudraVigilance), and patient reports. Data update frequencies vary; clinical trial data usually publishes after study completion, while adverse event reporting system data updates continuously in real-time or near real-time. Document structures are diverse, including structured database records, semi-structured XML or JSON reports, and unstructured PDF literature or free-text descriptions. Key fields include drug name, adverse event name (using MedDRA coding), occurrence date, patient demographics (e.g., age, gender), tumor type (e.g., acute myeloid leukemia AML, multiple myeloma MM), treatment regimen, event severity, outcome, and reporting source. Some data also contains genetic mutation information or biomarker data.
Constraints on Citation and Traceability
The multi-source and heterogeneous nature of hematologic oncology pharmacovigilance data imposes specific requirements on FastGPT's citation and traceability capabilities. First, the presence of specialized terminology like MedDRA codes requires the RAG retrieval model to accurately match and understand medical concepts, avoiding miscitations due to semantic ambiguity. Second, varying data update frequencies necessitate considering data timeliness during knowledge base construction. This ensures the latest version information is cited, especially for rapidly changing drug approvals and adverse event signals. Third, diverse document structures require flexible file parsing strategies. This effectively handles structured database records and unstructured text, ensuring all relevant information is extracted and indexed. Finally, detailed patient and disease characteristics, such as tumor type and genetic mutations, mean that citations must not only point to original literature or reports but, in some scenarios, also provide context specific to patient subgroups or particular treatment regimens to support rigorous drug risk assessment.
Configuration Guidelines
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
maxContext | 2000 characters | Accommodates key medical information such as MedDRA codes, drug names, tumor types, and adverse event descriptions, preventing truncation. |
Recall count | 8 | The complexity of hematologic oncology drug adverse reactions requires recalling information from multiple dimensions to increase coverage. |
Similarity threshold | 0.75 | Ensures high relevance between recall results and query intent, reducing inaccurate or generalized medical information citations. |
Chunk size | 500 characters | Balances the integrity of medical text and retrieval efficiency, preventing excessive fragmentation that leads to context loss. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Ensures sufficient time for file parsing when processing large PDF documents or reports with complex tables. |
Rerank result count | 5 | Focuses on the most critical citation sources through re-ranking while maintaining relevance, improving result usability. |
Common Pitfalls
- Missing or incomplete citation sources: Answers lack specific literature or report links. The file parser fails to correctly identify or extract citation metadata from unstructured documents.
- Generalized adverse event descriptions for specific tumor types (e.g., multiple myeloma): Results do not differentiate risk variations under different treatment regimens. The knowledge base construction did not adequately use fine-grained labels like disease subtypes and treatment regimens for indexing.
- Inability to reference variables for file upload parameters in custom workflows during tool calls: Uploading files requires manually providing links. The tool interface design restricts dynamic variables for file paths or content.
Validation Steps
- For specific hematologic oncology drug adverse event queries, check if the returned results include multiple clear citation sources. Verify accuracy and completeness by tracing back to original reports or literature.
- Input a query containing MedDRA codes. Verify FastGPT correctly understands and recalls precise medical information related to the code. Confirm it differentiates subtle differences between various codes.
- Simulate queries for adverse reactions under specific drugs for different tumor types (e.g., acute lymphoblastic leukemia vs. chronic myeloid leukemia). Observe if citation sources reflect disease specificity and evaluate thresholds based on actual conditions.
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.