Data Characteristics
Solid tumor pharmacovigilance data primarily comes from clinical trial reports, real-world evidence (RWE) studies, case reports, medical literature, and safety updates from drug regulatory agencies. This data often exists in a mixed format, combining structured records (e.g., database entries) and unstructured text (e.g., free-text descriptions). Data update frequency varies: clinical trial data is released periodically as studies progress, regulatory safety updates may be monthly or quarterly, and medical literature is continuously published. Document structure for clinical trial reports and post-market safety reports typically includes detailed patient characteristics, medication information, adverse event (AE) descriptions, severity, onset time, outcomes, and causality assessments. Fields include, but are not limited to: Patient ID, Drug Name, Adverse Event Term (MedDRA code), Date of Occurrence, Outcome, Causality Assessment. Units involved include dosage (mg, g), frequency (times/day), and duration (days, weeks, months).
Constraints on Source Citation and Traceability
The highly specialized and complex nature of solid tumor pharmacovigilance data imposes specific requirements on source citation and traceability. First, data source diversity necessitates support for ingesting and processing multi-format documents, such as PDF clinical trial reports and XML database export files. Second, adverse event descriptions often contain extensive medical terminology and abbreviations, requiring strong domain understanding from text segmentation and embedding models to prevent information fragmentation. The uncertain update frequency, especially for regulatory safety updates, demands a flexible knowledge base update mechanism to ensure citation timeliness. Finally, numerical information like dosage and time in reports requires precise referencing to the exact values and units in the original text, avoiding misinterpretation or ambiguity. This directly impacts RAG retrieval granularity and context window settings.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk Length | 500–700 characters | Solid tumor adverse event descriptions are often long with strong contextual relevance; sufficient information must be retained for causal inference. |
Recall Count | Top 8–12 entries | Ensures that under complex queries, relevant adverse event reports or medical literature snippets are recalled from multiple dimensions. |
Similarity Threshold | 0.75–0.82 | Balances retrieval precision and recall, avoiding interference from irrelevant medical terms or abbreviations. |
Reranked Count | Top 3 entries | Focuses on core evidence most directly related to solid tumor adverse events, reducing redundant information. |
maxContext | 3000–4000 tokens | Needs to accommodate key information from multiple adverse event reports and medical background to support complex reasoning. |
UPDATE_INTERVAL_DAYS | 7 days | Adapts to the frequency of regulatory safety updates and medical literature publications, maintaining knowledge base timeliness. |
Common Pitfalls
- Incomplete medical terms or numerical values appear in query results because the chunk length is set too short, truncating critical information.
- The system fails to recall the latest drug safety updates because no periodic knowledge base update task is configured or executed.
- The original passage pointed to by the citation source has weak relevance to the query result because the similarity threshold is set too low, introducing a large amount of noise data.
Validation Steps
- Select typical solid tumor adverse event queries. Check if the returned results include multiple relevant sources and verify that each source points precisely to the original passage.
- Periodically simulate the release of new drug safety announcements. After the knowledge base updates, check if relevant queries can accurately cite the new content.
- For queries containing numerical information like dosage and frequency, verify that the numerical values and units in the returned results exactly match the citation sources.
- Conduct multi-turn dialogue tests. Confirm that when asking for detailed information about adverse events, the system can trace back from historical citations and provide coherent answers.
The values given are common starting points and should be measured against the reader's 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.