Data Characteristics
Patient assistance program data originates from pharmaceutical companies and Contract Research Organizations (CROs). This data includes clinical trial protocols, investigator brochures, informed consent forms, and patient inclusion/exclusion criteria. Documents are typically in PDF, Word, or structured database formats. Data update frequency varies based on trial phases and protocol revisions, usually occurring after trial initiation, protocol amendments, or regulatory approvals. Document structures are complex, containing extensive medical terminology, numerical ranges, disease diagnostic criteria, and concomitant medication contraindications. Fields include disease names, gene mutation types, KPS scores, ECOG scores, specific biomarker test results, and prior treatment history. Units involve dosages (mg), time (weeks, months), and numerical ranges (e.g., blood routine indicators).
Constraints on Knowledge Base Retrieval and Recall
Complex medical terminology and synonym variations challenge recall accuracy, requiring vocabulary expansion and semantic understanding capabilities. Numerical ranges and multi-condition combination filtering demand a retrieval system capable of parsing and matching complex logical expressions; simple keyword matching can lead to omissions. Irregular document updates mean the knowledge base must support incremental updates and version management to ensure retrieval result timeliness. Key information scattered across long documents, such as inclusion/exclusion criteria, requires effective segmentation strategies to capture context and avoid information fragmentation. Heterogeneity of multi-source data, such as coexisting structured data and unstructured text, increases knowledge base construction complexity, necessitating unified indexing and query interfaces.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 800–1200 characters (characters) | Adapts to clinical trial document paragraph lengths, balancing contextual completeness and retrieval efficiency. |
Chunk Overlap Length (Segment Overlap Length) | 100–200 characters (characters) | Ensures key information correlation across paragraphs, preventing important information from being cut off at segment boundaries. |
Recall count (Number of Retrieved Items) | Top 8–12 entries (top 8–12 items) | Reduces the burden on subsequent model processing while ensuring coverage, avoiding interference from low-relevance results. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Balances recall rate and accuracy, reducing the risk of false positives for irrelevant documents, especially for medical terminology. |
Rerank result count (Number of Reranked Items) | Top 5 entries (top 5 items) | Further refines retrieval results, prioritizing the most relevant document snippets to enhance user experience. |
Embedding Model Version | text-embedding-ada-002 | Widely used in the industry, with good semantic understanding capabilities for medical texts. |
Common Pitfalls
- The number of query results is unusually low, or expected documents are not retrieved. This occurs due to improper segmentation strategies, where key information is split across different segments, or the embedding model's understanding of medical terminology is insufficient.
- When faced with queries containing numerical ranges (e.g., "KPS score greater than 70"), the system returns irrelevant results. This happens because the knowledge base does not specially process numerical or logical queries, performing only text matching.
- After a knowledge base update, query results still contain old information. This is due to an incomplete incremental update mechanism or untimely index reconstruction, leading to retrieval of outdated data.
Verification of Configuration
- Develop a standard query set for typical patient profiles and inclusion/exclusion criteria. Verify that retrieval results include all necessary document snippets.
- Design complex queries involving numerical ranges and multi-condition combinations. Check if retrieval results accurately filter documents that meet the logical criteria.
- Simulate the knowledge base update process. Observe the timeliness and accuracy of new and old information retrieval to ensure the system responds quickly to data changes.
Note: The values provided are common starting points. Measure performance against your own samples to determine optimal configurations.
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.