Data Characteristics
Solid tumor quality documents primarily originate from regulatory agency guidelines, clinical trial protocols, investigator brochures, drug inserts, internal Standard Operating Procedures (SOPs), and batch production records. These documents update infrequently, typically following regulatory release cycles or internal version control processes, such as annual or multi-year revisions. Document structures are mostly unstructured text, often including numerous tables, figures, and attachments. The text content is highly specialized, covering tumor typing, treatment regimens, drug dosages, side effects, and clinical endpoints. Fields and units are highly specific; for example, tumor size often uses mm or cm, and drug dosages use mg/kg or mg, frequently accompanied by complex medical terminology and abbreviations.
Constraints on Knowledge Base Retrieval and Recall
The low update frequency of solid tumor quality documents means that, after knowledge base construction, the update strategy can focus on version management, reducing frequent full re-indexing. The presence of unstructured text and numerous figures demands higher document parsing capabilities to ensure accurate text extraction. Dense specialized terminology and abbreviations can lead to insufficient recall rates with traditional keyword-based retrieval, requiring enhanced semantic understanding and conceptual association. The existence of specific fields and units requires the retrieval system to identify and match this key information; for example, when querying a specific dosage range, the system must understand the context of mg/kg or mg. These constraints collectively point to the need for optimizing knowledge base chunking strategies and retrieval model accuracy.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk Length | 500–800 characters | Ensures individual chunks contain sufficient contextual information while avoiding excessive length that could lead to redundancy and semantic drift. |
Chunk Overlap | 50–100 characters | Maintains contextual continuity and reduces the loss of critical information due to chunk boundary cuts. |
Retrieval Count | 8–12 items | Balances retrieval coverage with avoiding excessive irrelevant information that could impact subsequent re-ranking efficiency. |
Similarity Threshold | 0.75–0.85 | Balances accuracy and recall, filtering for highly relevant document snippets and reducing noise. |
Rerank Return Count | 3–5 items | Focuses on the most relevant core information, improving the precision of the final output and user experience. |
Max Context Window | 4096 tokens | Accommodates the complex descriptions and detailed information typical of solid tumor documents, ensuring complete understanding. |
Common Pitfalls
- Retrieval results contain numerous irrelevant or outdated clinical trial details. This happens when the knowledge base lacks effective metadata filtering to distinguish formally published quality documents from research materials.
- Queries for specific drug dosage ranges return no results. This occurs when the chunking strategy fails to effectively preserve or identify numerical fields with units.
- User questions about the latest treatment plans for a specific tumor return old guideline versions. This happens when the knowledge base update mechanism is not synchronized with regulatory document release processes, leading to a failure to index the latest document versions promptly.
How to Verify Configuration
- Select a batch of test questions covering different tumor types, treatment plans, and drug dosages. Observe whether retrieval results cover all key information points and compare them against expected outcomes to confirm retrieval accuracy.
- Construct queries using specific medical terminology and abbreviations found in the documents. Check if the system correctly identifies and retrieves corresponding explanations or context from relevant document snippets, such as querying the mechanism of action for
PD-1inhibitors. - Simulate updates to different versions of quality documents. Perform queries after uploading new versions to verify if the knowledge base prioritizes the latest relevant content and if the retrieval priority for older version information decreases.
Note: The values provided are common starting points and should be measured against your 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.