Data Characteristics
Respiratory system pharmacovigilance data comes from various sources. These include drug labels, clinical study reports, real-world evidence (RWE) databases, adverse event reporting systems (e.g., MedDRA), and academic journal literature. Data updates frequently, especially post-market surveillance reports and new clinical guidelines, typically quarterly or annually. Document structures are primarily semi-structured and unstructured. For example, adverse event reports often contain free-text descriptions, coded fields, and timestamps. Key fields include specific side effect descriptions, mechanisms of action, drug interactions, patient baseline characteristics (e.g., age, underlying diseases), and dosage with adverse event incidence rates. Units for dosage are often milligrams (mg) or micrograms (mcg). Time units include days, weeks, and months. Adverse event incidence rates are expressed as percentages or per thousand person-years.
Constraints on Knowledge Base Retrieval and Recall
The diversity and update frequency of respiratory system pharmacovigilance data challenge knowledge base recall efficiency and accuracy. Semi-structured and unstructured documents require fine-grained text processing and entity recognition to ensure no critical information is missed. High update frequency means the knowledge base must support incremental updates and version management to avoid retrieving outdated information. Medical terminology and synonym variations in specific side effect descriptions demand strong semantic understanding from the retrieval model to identify similar concepts expressed differently. The presence of numerical fields like dosage and time requires retrieval to match text and perform numerical range queries or filtering based on numerical context. The large and complex data volume can increase retrieval response times, requiring high system performance, especially for multi-field, multi-condition complex queries.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size | 400–600 characters | Respiratory adverse event reports often contain long free-text descriptions. This length helps preserve context and reduces information fragmentation. |
Chunk Overlap Length | 50–100 characters | Ensures sufficient contextual overlap between adjacent segments, improving semantic coherence across segments. |
Recall count | 8–12 entries | Considering the complexity and potential associations of respiratory drug adverse reactions, increasing the number of recalled items helps cover more potentially relevant information. |
Similarity threshold | 0.7–0.8 | Ensures high relevance between retrieval results and the query, filtering out irrelevant or weakly related documents, and reducing noise. |
Rerank result count | 3–5 entries | Reranking after initial recall selects the most relevant results, improving the quality and efficiency of information presented to the user. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Extending the file parsing timeout when processing large clinical reports or complex multi-page PDF documents prevents parsing failures due to large file sizes. |
Common Mistakes
- Symptom: Knowledge base retrieval returns empty results or results clearly inconsistent with the query. Reason: Medical terminology standardization and synonym processing were insufficient during data preprocessing. This prevents query terms from matching actual content in the knowledge base.
- Symptom: Knowledge base search nodes execute for too long in the workflow, averaging over 5 seconds. Reason: The knowledge base index is not optimized, or the text segmentation strategy is too granular. This leads to processing many small fragments during retrieval, increasing computational burden.
- Symptom: Queries for specific drug dosages or time ranges return inaccurate results. Reason: Numerical information was not correctly extracted as queryable metadata fields during knowledge base construction, or the retrieval model lacks support for numerical range queries.
How to Confirm Correct Configuration
- For typical respiratory adverse drug reaction queries, verify that retrieval results include key drug, adverse event, dosage, and time information.
- Use queries containing medical synonyms and abbreviations. Check if the knowledge base correctly identifies and recalls relevant documents. Compare these results with those obtained before standardization.
- Randomly select a batch of recently updated literature or adverse event reports. Execute queries and confirm the knowledge base can promptly recall new data. Update timeliness should meet business requirements.
- For specific drugs, design queries including different dosages and time windows. Check if retrieval results accurately reflect numerical contextual relationships.
Note: The values provided are common starting points. Measure them against your 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.