Data Characteristics
Hematologic oncology R&D documents originate from clinical trial reports, pathology analysis reports, gene sequencing data, drug mechanism of action research papers, and regulatory submissions. This data updates frequently, especially clinical trial interim reports and new research findings, typically on a monthly or quarterly basis. Document structures are highly complex, containing extensive specialized terminology, abbreviations, and specific medical charts. Field specificity is high, including terms like "Complete Remission Rate (CR%)", "Minimal Residual Disease (MRD) status", and "Chromosomal Karyotype Abnormalities". Units are diverse, covering concentrations (nM, µg/mL), dosages (mg/kg), and time (days, months, years). Some values are unitless, from specific disease scoring systems. Pathology reports often include images and tables, requiring multimodal analysis.
Constraints on Reference and Traceability
The complexity of hematologic oncology R&D documents imposes strict requirements on reference and traceability. High-frequency data updates necessitate efficient incremental updates and version management for the knowledge base. This ensures references always point to the latest valid evidence. Complex document structures and dense specialized terminology require precise identification and extraction of key information during structured analysis. Examples include drug names, targets, clinical endpoints, and corresponding values. Diverse fields and units demand accurate parsing and standardization to prevent traceability errors due to unit confusion. The presence of pathology images and tables requires multimodal processing. This ensures all relevant information is included in retrieval and referencing, allowing users to precisely trace back to specific charts in reports during conversations. Accurate identification and traceability of reference sources directly impact the scientific rigor and safety of R&D decisions.
Configuration Guidelines
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 400–600 characters | Ensures completeness of key information (e.g., an experimental result, a patient cohort description), preventing semantic fragmentation. |
Recall count (Recall Count) | 8–12 items | Covers multiple potentially relevant document segments, increasing recall comprehensiveness, especially for complex queries. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Balances recall accuracy and completeness, reducing interference from irrelevant medical terms. |
Rerank result count (Reranked Return Count) | 5 items | Focuses on the top most relevant results, reducing the large language model's burden of processing irrelevant information. |
maxContext | 4000 tokens | Accommodates the information density in hematologic oncology documents, allowing the model to process longer contexts for reasoning. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Handles parsing of large clinical trial reports or pathology reports with numerous charts, preventing timeouts. |
Common Pitfalls
- Symptom: AI responses cite content that does not fully match the disease type or drug target in the user's query. Reason: The
Similarity threshold(Similarity Threshold) is set too low, leading to the recall of semantically imprecise document segments. - Symptom: A specific experimental data point mentioned in the conversation cannot be traced back to its exact location in the original document. Reason: The
Chunk size(Chunk Size) is set too long. A single chunk contains too much irrelevant information, preventing the model from focusing on the most critical statements when citing. - Symptom: After a knowledge base tool call in the workflow, the model cannot cite key data from it. Reason: The knowledge base query traffic is too high, or
maxContextis set too low. This prevents the large language model from fully utilizing all returned knowledge when processing retrieval results.
Validation Steps
- For a specific hematologic oncology disease or drug, submit a series of complex questions containing specialized terminology. Check if the AI's response precisely points to the specific paragraphs in the original document.
- Upload a pathology report containing tables and images. Ask about key data within it. Verify if the AI can correctly cite and trace back to information within the images or tables.
- After a knowledge base update, immediately test relevant questions. Confirm the AI can cite the latest version of the data. Verify the timeliness of the citation by comparing it with the original document.
- Simulate knowledge base queries under high concurrency. Monitor system response times. Check if the
PARSE_FILE_TIMEOUT_SECONDSsetting is sufficient to handle the parsing of the longest documents.
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.