Data Characteristics in Hematologic Oncology
R&D documents in hematologic oncology typically include clinical trial reports, pathological analysis reports, gene sequencing data, drug mechanism of action studies, and various experimental records. These documents originate from diverse sources, such as internal pharmaceutical company databases, public medical literature, and disease center reports. Data updates frequently, especially for new drug development progress and clinical trial results, with new additions possibly weekly or even daily. Document structures are complex, containing both highly structured tabular data and extensive unstructured text descriptions, such as adverse event records and patient follow-up information. Fields and units are highly specialized, for example, "Complete Remission Rate (CR%)", "Overall Survival (OS)", gene mutation sites (e.g., "JAK2 V617F"), drug dosages (e.g., "mg/kg"), and cell counts (e.g., "cells/μL"). These specific terms and units demand high accuracy in analysis.
Constraints Imposed by These Characteristics on Model Access and Configuration
The complex data structure of hematologic oncology R&D documents requires models with multimodal processing capabilities to understand text, tables, and charts simultaneously. High-frequency data updates mean the knowledge base needs efficient incremental update mechanisms to avoid frequent full rebuilds. The extensive use of specialized terminology, abbreviations, and disease-specific concepts (e.g., myelodysplastic syndromes, lymphoma) in documents requires models to learn relevant domain knowledge during pre-training or fine-tuning. Failing to do so can lead to anaphoric resolution errors or semantic understanding deviations. Furthermore, diverse fields and units demand strict accuracy from the model in information extraction and entity recognition, necessitating targeted configuration of entity recognition models or dictionaries to ensure correct association of values and units.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size | 800–1200 characters | Balances semantic completeness and recall efficiency, prevents context loss, and controls token consumption. |
Chunk Overlap Length | 100–200 characters | Ensures semantic continuity between paragraphs and reduces the risk of cutting key information. |
Recall count | Top 5–8 entries | Considers information density and model processing capacity to cover main relevant content. |
Similarity threshold | 0.75–0.85 | Filters out low-relevance results to avoid noise; adjust based on actual data. |
Rerank result count | Top 3 entries | Further refines recall results to improve the precision of the final answer. |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Accommodates the time required to process large clinical trial reports or gene sequencing data, preventing timeouts. |
Three Common Mistakes
- During document parsing, a "missing parameter or incorrect format" error appears. This happens when some specialized charts or nested table structures are too complex for the default parser to correctly identify key fields.
- The model's answer shows incorrect understanding of drug dosages or gene loci, with mismatched values and units. This occurs due to the lack of a domain-specific vocabulary configured for hematologic oncology terminology and units.
- Knowledge base query results do not meet expectations, for example, insufficient recall of literature on a specific disease (e.g., multiple myeloma). This is because the base model fails to adequately expand disease-related synonyms and specialized terms during query rewriting.
How to Confirm Proper Configuration
- Upload typical hematologic oncology R&D documents (e.g., a multiple myeloma clinical trial report) and check if the document content is fully parsed, especially key tabular data and gene sequence information.
- Perform knowledge base queries for specific specialized terms (e.g., "BCMA-targeted therapy", "CAR-T cell therapy") to evaluate the model's understanding and contextual association capabilities for these terms.
- Submit questions containing complex references and abbreviations (e.g., "FLT3 mutation in AML patients") to observe if the model's pre-retrieval query rewriting is accurate and confirm if the final answer correctly points to the original document.
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.