Data Characteristics
Medical Information (MI) response data primarily originates from pharmaceutical companies' internal medical information databases, clinical trial reports, drug monographs, medical journal literature, and industry conference materials. This data updates frequently. New drug approvals, clinical research advancements, and adverse drug reaction reports trigger data updates. Systemic maintenance typically occurs quarterly or semi-annually, but urgent safety information updates can happen at any time. Data document structures vary, including structured database records, semi-structured PDF documents, Word reports, and unstructured text content. Beyond common fields like drug name, indication, and dosage, data also includes extensive medical terminology, disease codes (e.g., ICD-10), drug interactions, clinical study data (e.g., P-values, CI values), and various units of measurement (e.g., mg/kg, IU, mmol/L).
Constraints on Citation and Traceability
The complexity and high update frequency of MI response data sources impose strict requirements on citation accuracy and traceability. First, diverse document structures mean a RAG system needs robust multi-format parsing capabilities to ensure correct extraction and indexing of all relevant information. Second, accurate identification of specialized medical terminology and units of measurement directly impacts the accuracy of cited content. Any ambiguity could lead to serious medical risks. High update frequency necessitates timestamp management for citations, ensuring the latest information version is cited and avoiding outdated or corrected data. Furthermore, since MI responses often support healthcare professional decisions, citations must clearly trace back to original literature or database entries, allowing users to verify information reliability.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 3000 Tokens | Ensures sufficient context for complex medical concepts, preventing information truncation. |
Recall count | Top 10 entries | Increases the probability of recalling relevant medical literature or data entries, covering a broader range of information. |
Similarity threshold | 0.78 | Improves recall precision, filtering for medical content highly relevant to the query. |
Rerank result count | Top 5 entries | Selects the most relevant and authoritative medical information as final citation sources, reducing redundancy. |
Chunk size | 400 characters | Balances segment completeness and recall efficiency, ensuring medical terminology context remains intact. |
chunk_overlap_ratio | 0.1 | Maintains contextual continuity between segments, facilitating understanding of complex medical concepts. |
Common Pitfalls
- The number of context items displayed on the page does not match the actual number sent to the model. This often occurs due to discrepancies between front-end display logic and back-end processing logic, or unreflected data compression/filtering in the front-end.
- Citations contain a large amount of non-medical content. This indicates an improper knowledge base segmentation strategy or a similarity threshold set too low, failing to effectively filter irrelevant information.
- The model's response cites outdated or corrected medical information. This suggests the knowledge base update mechanism is not synchronized with the data source's update frequency, leading to stale indexed data.
Verification Steps
- Randomly select 10 typical MI questions. Check if the model's response accurately points to original medical literature or database entries and verify consistency between cited content and the original text.
- For recently updated drug information or clinical guidelines, submit relevant queries. Verify if the model's citations are the latest version and check citation timestamps.
- Compare medical terminology and units of measurement in the model's response to ensure exact consistency with the original source, paying close attention to special symbols and numerical precision.
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.