Data Characteristics in This Domain
Rare disease data primarily originates from post-market surveillance reports, clinical trial reports, real-world evidence (RWE) data, medical literature, and patient registries published by regulatory bodies such as the National Medical Products Administration (NMPA), FDA, and EMA. Data update frequencies vary; regulatory reports are typically quarterly or annual, while medical literature updates continuously. Document structures are predominantly unstructured text, including case reports, research papers, and news articles. Semistructured data also exists, such as drug inserts and adverse event report forms (CIOMS I or MedWatch 3500A). Fields cover patient demographics, disease diagnosis, medication details (dosage, duration), adverse event descriptions, severity, and outcomes. Units include dosage units (mg, g), frequency units (times/day, times/week), and time units (hours, days, years).
Constraints Imposed by These Characteristics on "Citation and Traceability"
The dispersed nature and inconsistent update frequency of rare disease data necessitate a focus on multi-source data integration and version management during knowledge base construction. The high proportion of unstructured text demands higher quality RAG text segmentation to ensure segments fully capture the context of adverse event descriptions. Fields within semistructured data, such as drug dosage and adverse event severity, require precise identification and extraction to provide accurate quantitative information in citations. Due to the small number of rare disease patients, a single adverse event report may hold unique value. This requires the citation and traceability mechanism to precisely point to the original report, preventing information loss or generalization. Data may also contain sensitive patient information, requiring anonymization and adherence to data privacy regulations during citation and display. Discussions of specific adverse reactions often require aggregating fragmented information from multiple sources, which demands capabilities for aggregating and de-duplicating citation sources.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 300-500 characters | Balances the completeness of rare disease adverse event descriptions with retrieval efficiency. |
Overlap Length | 50 characters | Ensures contextual continuity between adjacent segments, reducing information fragmentation. |
Recall count (Recall Count) | 8-12 entries | Balances recall breadth with large model token limits, preventing omission of critical information. |
Similarity threshold (Similarity Threshold) | 0.75-0.85 | Filters out low-relevance documents, improves recall precision, and addresses data sparsity. |
Max Citation Tokens | 1500 token | Prevents large model context overflow due to overly long citations while providing sufficient information for traceability. |
Citation Display Mode | Title and Link Only | Ensures traceability visibility without occupying excessive answer space, allowing users to easily click and review original reports. |
Common Pitfalls
- The number of documents cited in the answer significantly exceeds expectations, leading to verbose responses and high token consumption. This occurs when
Recall count(Recall Count) orMax Citation Tokensare configured too high, failing to effectively limit the citation scope. - The answer includes cited content unrelated to rare disease pharmacovigilance. This may be due to a
Similarity threshold(Similarity Threshold) set too low, leading to the recall of significant noise. - Cited source links point to incorrect or inaccessible locations. This typically results from incorrect maintenance of source file paths or the
source_urlmetadata field during knowledge base import, or changes to the original data source.
How to Confirm Proper Configuration
- Randomly select 5-10 rare disease adverse reaction-related queries. Check if the cited documents in the answer are accurate and highly relevant to the question. Record the number of documents recalled.
- Verify that all cited source links in the answer are accessible and can pinpoint the original information.
- Assess whether the cited content in the answer comprehensively covers critical information from the original document, especially detailed descriptions of rare disease adverse events, dosages, and occurrence times.
The values provided 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.