Vector Models and Indexing for Hematologic Oncology

Data Characteristics in This Category

Registration documents in hematologic oncology have specific data characteristics. Data sources are extensive, including clinical trial reports, non-clinical study reports, pharmaceutical research data, regulatory guidelines, expert consensuses, and package inserts for marketed drugs. These documents have varying update frequencies. Regulatory guidelines and expert consensuses may update annually or irregularly, while clinical trial data is generated and aggregated in real-time as research progresses. Document structures are complex, often containing numerous tables, figures, and extensive text descriptions with dense specialized terminology. Field and unit specificities include medical indicators, dosage units, and statistical indicators. Examples are platelet count (PLT, unit: 10^9/L), overall response rate (ORR, unit: %), and progression-free survival (PFS, unit: months or days). Documents also contain many abbreviations, synonyms, and cross-document references.

Constraints Imposed by These Characteristics on Vector Models and Indexing

The complexity of hematologic oncology registration documents places specific demands on vector models and indexing. Long texts and dense specialized terminology require selecting vector models that effectively handle long contexts and understand medical semantics, preventing information loss or semantic deviation. Multi-source heterogeneous data requires indexing strategies that integrate information from different sources and formats, establishing effective connections. For example, adverse events in clinical trial reports and drug metabolism pathways in pharmaceutical data may need to be indexed together. Varying data update frequencies require the index to support incremental updates, ensuring the knowledge base's timeliness, especially for changes in regulations and guidelines. Specificity of fields and units, such as dosage and biomarkers, requires the model to distinguish between values and units and understand their medical meaning, avoiding vectorizing values as independent semantic units. Additionally, the presence of numerous abbreviations and synonyms requires vector models to have some word sense disambiguation and generalization capabilities.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk size` (Segment Length)	`800–1200 characters`	Balances semantic completeness of long texts and vector model processing capacity, avoiding excessive truncation of key information.
`Chunk Overlap Length` (Segment Overlap Length)	`150–200 characters`	Ensures contextual continuity and captures semantic connections across segments.
`embeddingModel`	`text-embedding-3-large` or `bce-embedding`	High-dimensional models better capture complex semantics in hematologic oncology and support Chinese medical texts.
`maxToken`	`8192`	Accommodates lengthy clinical trial reports and pharmaceutical data, preventing context truncation.
`Recall count` (Recall Count)	`8–12 entries`	Ensures sufficient relevant context is recalled from vast amounts of data, improving accuracy.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Balances recall and precision, filtering out irrelevant documents and focusing on core medical information.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Ensures sufficient time for parsing large PDF or Word documents.

Three Common Mistakes

Knowledge base query response times are too long, or timeout errors occur. This happens when insufficient parsing timeout is configured for large documents or an underperforming vector model is used.
Retrieval results contain many irrelevant regulatory documents or generalized information, leading to answers deviating from core medical questions. This is due to a Similarity threshold (Similarity Threshold) set too low, failing to effectively filter noise.
Queries for specific tumor types (e.g., leukemia) fail to recall all relevant clinical trial data. This may be because the Chunk size (Segment Length) is too short, causing key information to be fragmented or context lost.

How to Confirm Proper Configuration

Upload typical hematologic oncology registration documents (e.g., a complete clinical trial report). Check if document segmentation is reasonable and semantic integrity is maintained.
Query specific medical terms or clinical indicators. Verify the similarity scores of recalled documents and evaluate their relevance to the query, confirming if the threshold is set appropriately.
Simulate complex problems encountered during actual registration document preparation. Observe if the model's generated answers accurately cite multiple sources and include correct fields and units. This helps determine if Recall count (Recall Count) and maxToken are sufficient to support complex reasoning.

The values given 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.