Data Characteristics
Attenuated live vaccine pharmacovigilance data originates from national drug adverse reaction monitoring centers, vaccine manufacturer reporting systems, spontaneous reports from medical institutions, and scientific literature. Data updates typically occur quarterly or annually, but reports of major events may be real-time. Document structures are primarily structured reports, such as the "Adverse Drug Reaction Report Form," which includes fields for patient basic information, medication history, adverse reaction description, and outcome. Unstructured data, such as clinical medical records and follow-up notes, are also common. Key fields include vaccine batch number, administration route, adverse reaction onset time, severity level (e.g., CTCAE v5.0 standard), and medical terminology codes (e.g., MedDRA codes). Units adhere to international standards, such as mg/mL for dosage and hours/days for time.
Constraints on Citation and Traceability
The diverse sources of attenuated live vaccine pharmacovigilance data require citation sources to integrate different data formats, ensuring comprehensive knowledge base coverage. Varying update frequencies necessitate different data synchronization strategies. For example, regulatory agency reporting systems may use periodic full synchronization, while manufacturer systems may use incremental updates. The coexistence of structured reports and unstructured text means that knowledge base document segmentation must consider both field semantics and contextual integrity to avoid fragmenting critical information. The use of specialized terminology like MedDRA codes requires domain adaptation for text vectorization models to ensure accurate similarity calculations. Precise matching of key identifiers, such as batch numbers, is fundamental for effective traceability and should be prioritized in citation logic.
Configuration Guidelines
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size | 800–1200 characters | Balances the completeness of structured report fields with the contextual semantics of unstructured text. |
Recall count | Top 8 entries | Ensures coverage of multi-source data and improves recall rate for relevance. |
Similarity threshold | 0.78–0.85 | Precision requirements for vaccine adverse reaction descriptions to avoid interference from irrelevant information. |
Rerank result count | Top 3 entries | Prioritizes the most relevant and authoritative citations, reducing user reading burden. |
Parsing Strategy | Combination of intelligent parsing and custom rules | Adapts to mixed data formats such as structured reports and clinical medical records. |
maxContext | 4096 tokens | Ensures sufficient original citation text for the model to perform in-depth analysis. |
Common Pitfalls
- The model's answer does not cite local knowledge base content, but the citation list includes relevant entries. This may occur if knowledge base segments are too fine-grained or the similarity threshold is set too high, leading to retrieved segments that do not sufficiently match the user's query semantics for the model to adopt as the main answer.
- The Chat interface returns the citation list before the main answer. This typically indicates an asynchronous processing mechanism, where model answer generation takes time, while citation retrieval and display may complete sooner.
- After knowledge base content updates, the model still cites old data. This may be due to the knowledge base index not being rebuilt promptly or the cache not being refreshed, causing the model to still recall based on old vector indexes.
Validation Steps
- For typical attenuated live vaccine adverse reaction queries, check if the model's answer includes correct vaccine batch numbers, adverse reaction types, and reporting sources, and verify that the citation list points to the accurate location in the original document.
- Test adverse reaction queries of different severity levels and outcomes to verify the comprehensiveness of citation sources, ensuring coverage of various data types (e.g., post-market surveillance reports, clinical trial data).
- Simulate data update scenarios and observe the timeliness of the model's citation of new data after a knowledge base update, by querying newly entered adverse reaction events.
- Check the accuracy of the model's citation of specialized terminology like MedDRA codes in its answers, confirming the knowledge base's parsing and indexing capabilities for professional terms.
Note: 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.