Data Characteristics for This Category
Real-World Study (RWS) product data originates from Electronic Health Records (EHR), medical insurance claims data, patient registries, wearable devices, and Patient-Reported Outcomes (PROs). This data typically exists as unstructured text (e.g., clinical notes, pathology reports), semi-structured data (e.g., lab results, medication records), and structured data (e.g., diagnosis codes ICD-10, drug codes NDC). Data update frequencies vary; some sources like EHRs update in real-time, while insurance claims data may have a delay of weeks to months. Document structures are complex, containing extensive medical terminology, abbreviations, and specific formats. Fields and units are diverse, including dosage units (mg, µg), time units (days, months, years), and various biomarker units.
Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall
The highly heterogeneous and unstructured nature of RWS data presents challenges for knowledge base text segmentation and vectorization. The specialized nature of medical terminology requires vector models with domain knowledge to accurately capture semantic similarity. Data update lag, especially for studies involving long-term follow-up, means the knowledge base must support incremental updates and version management to avoid recalling outdated information. Structured and semi-structured information within documents, such as diagnosis codes and drug dosages, requires preprocessing or indexing as metadata to support more precise filtering and retrieval. The diversity of fields and units necessitates standardization or normalization during text processing to reduce retrieval bias caused by inconsistent units.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 500–800 characters | Balances context completeness for medical text with vector model processing efficiency, preventing information overload or context loss in a single chunk. |
Chunk overlap (Chunk Overlap) | 100 characters | Ensures critical information spanning across segments is not cut off, improving retrieval recall rate. |
Recall count (Recall Count) | 10–15 items | Considering the complexity of RWS queries, increasing recall quantity appropriately covers more potentially relevant information. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Balances recall and precision, avoiding the retrieval of numerous irrelevant but lexically similar medical terms. |
Rerank result count (Reranked Return Count) | 3–5 items | After selection by a reranking model, provides the most relevant core information, reducing the processing burden on downstream LLMs. |
Text Understanding Model | Domain-fine-tuned model | Optimized for specialized terminology and context in the biomedical field, improving the accuracy of vector representations. |
Three Common Pitfalls
- Recall results have low relevance to the question, with generated text appearing in the answer that is unrelated to the knowledge base content. This occurs when the
Similarity threshold(Similarity Threshold) is set too low, leading to the recall of many irrelevant documents, or when a reranking model is not enabled. - Retrieval results include outdated or deprecated research protocols. This occurs when the knowledge base is not regularly updated or lacks a version management mechanism, causing the model to retrieve old data.
- Retrieval results are inaccurate for queries containing specific drug dosages or diagnosis codes. This occurs when structured information is not indexed as metadata, or when key numerical and unit information is not preserved during text segmentation.
How to Confirm Proper Configuration
- For typical RWS queries, examine the recalled raw document snippets to confirm they contain query keywords and relevant context.
- Verify the reranked document snippets, ensuring they are highly semantically relevant to the query and directly support the generation of the final answer.
- Check the knowledge base update logs to confirm that the latest research data has been successfully imported and indexed, ensuring data timeliness meets requirements.
- Evaluate the retrieval system's ability to accurately identify and recall relevant information by simulating queries containing specific codes (e.g.,
ICD-10code U07.1) or dosage units (e.g.,50 mg/kg).
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.