Data Characteristics in This Category
Respiratory system pharmacovigilance data primarily originates from national drug adverse reaction monitoring centers, pharmaceutical manufacturers, medical institutions, and spontaneous patient reports. Data updates frequently. Adverse reaction reports are typically submitted within hours to days; serious adverse reaction reports have strict time limits. Document structures vary, including standardized CIOMS I forms, MedWatch forms, clinical trial reports, and unstructured case descriptions and medical literature. Fields cover patient basic information, drug information (e.g., generic name, brand name, batch number, dosage form, usage and dosage), adverse reaction descriptions (including medical terminology and natural language descriptions), prognosis, and related medical events. Units commonly include milligrams (mg) and micrograms (μg) for dosage, days and hours for time, and times/day for frequency.
Constraints Imposed by These Characteristics on "Citation and Traceability"
High-frequency updates and diverse data sources demand real-time and comprehensive citations. This ensures AI responses reflect the latest pharmacovigilance information. Unstructured text, especially patient spontaneous reports and medical literature, requires the knowledge base to have strong text parsing and semantic understanding capabilities. This accurately extracts adverse events and related drugs. Original documents stored across different platforms lead to a poor user experience when directly citing. Embed original links as traceability information in AI responses. Different sources may contain redundant or conflicting information. Limiting citation count and setting similarity thresholds is crucial to avoid presenting excessive or inconsistent citations. Accurate matching and identification of specific fields (e.g., drug batch numbers, adverse reaction medical terms) are key to ensuring traceability accuracy.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Recall count | Top 5-8 entries | Balances information comprehensiveness with response conciseness, avoiding redundant citations. |
Similarity threshold | 0.75-0.85 | Ensures relevance of recalled content, filtering out irrelevant or weakly related information. |
Chunk size | 400-600 characters | Accommodates the common medium-to-long length descriptions in respiratory adverse reaction reports. |
Rerank result count | Top 3 entries | Focuses on the most relevant and authoritative citation sources, improving traceability efficiency. |
Reference Link Template | https://original_source.com?doc_id={doc_id} | Ensures AI responses can directly provide clickable links to original documents. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Addresses potentially long parsing times for large clinical trial reports or literature. |
Three Common Mistakes
- Quoted original document links in AI responses are broken, displaying a 404 error. This occurs when the knowledge base fails to correctly parse or store the permanent link of the original document during import, or the link template variable
doc_idis not correctly populated. - AI responses show multiple citation sources, but content is contradictory or repetitive. This may be due to a
Similarity thresholdset too low, recalling many weakly related or overlapping document segments, or failing to effectively deduplicate recall results. - For adverse reaction queries targeting specific drug batch numbers, the AI response fails to cite reports containing that batch number. This may occur if the knowledge base chunking strategy is too coarse, separating batch number information from critical adverse reaction descriptions, affecting recall accuracy.
How to Confirm Correct Configuration
- Randomly select 10-15 respiratory drug adverse reaction queries. Check if all original links cited in AI responses are accessible and point to the correct document content.
- For responses with multiple citation sources, manually compare the content of each citation. Confirm information consistency and assess for redundancy.
- Select queries involving specific drug batch numbers, serious adverse events, or rare adverse reactions. Verify if the AI response accurately cites document segments containing this key information and traces back to the original report.
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.