Data Characteristics in This Category
Solid tumor-related regulations and SOP documents in the biopharmaceutical field typically originate from the National Medical Products Administration (NMPA), various medical institutions, and industry associations. These include clinical practice guidelines, drug inserts, ethical review documents, and clinical trial protocols. Document update frequencies vary: regulatory files might update annually, clinical guidelines revise periodically based on research advancements, and internal hospital SOPs may change less often. Most documents are in PDF format, containing numerous tables, images, appendices, and multi-level headings. The text is rigorous, with dense professional terminology, including dosage units (e.g., mg/kg), time units (e.g., weeks, months), biomarker names, gene mutation sites, and often includes reference lists.
Constraints Imposed by These Characteristics on Citation and Traceability
The rigor of solid tumor regulatory documents demands highly accurate and traceable Q&A results, especially for critical information such as dosage and treatment duration. Complex tables and images in documents mean that text-only retrieval and citation might miss key information. Multi-level headings and appendices require retrieval results to precisely locate specific sections or page numbers in the original text, allowing users to quickly verify information. Varying update frequencies necessitate a robust knowledge base update mechanism to ensure citations are from the latest versions. The dense presence of professional terminology and units means the model needs high semantic recognition capabilities during understanding and recall to avoid citation deviations due to synonyms or near-synonyms.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 500–800 characters | Ensures each knowledge chunk contains sufficient context while avoiding excessive length, which could lead to information redundancy and reduced recall efficiency. This also helps maintain the integrity of tables and paragraphs. |
Recall count (Recall Count) | Top 8–12 items | Given the complexity and multi-dimensional nature of solid tumor knowledge, increasing the recall count covers more potentially relevant information, improving coverage. |
Similarity threshold (Similarity Threshold) | Calibrated by actual measurement | This needs to ensure high-relevance recall while avoiding omissions due to a threshold that is too high, or noise introduction from a threshold that is too low. |
Rerank result count (Reranked Return Count) | Top 5 items | Combined with the rigor of biopharmaceutical content, a reranking mechanism further optimizes the order of recall results, improving answer accuracy. |
maxContext | 3000–4000 tokens | Provides the model with an ample context window to process complex logical relationships and multi-chapter citations within solid tumor regulations. |
CHUNK_OVERLAP_SIZE | 50–100 characters | Appropriate chunk overlap helps maintain semantic coherence across chunks, especially for citations spanning pages or paragraphs. |
Three Common Mistakes
- Phenomenon: The AI answer does not cite knowledge base content or the cited content has low relevance to the question. Reason: The
Similarity threshold(Similarity Threshold) is set too high, filtering out slightly less relevant but still useful knowledge chunks. - Phenomenon: The answer includes outdated clinical guidelines or regulatory information. Reason: Knowledge base documents are not updated promptly, or a document version management mechanism is missing, causing the model to cite old data.
- Phenomenon: Citation sources point to incorrect sections or paragraphs, preventing users from quickly verifying information. Reason: Document chunking granularity is too large, or the indexing model fails to effectively recognize multi-level headings and table structures.
How to Confirm Proper Configuration
- Select a batch of solid tumor questions involving dosages, treatment courses, and specific biomarker diagnostic criteria. Check if the AI answer precisely cites specific values and units from the original text and can locate the correct section.
- Upload a document containing both new and old versions of regulations. Ask questions about policy changes and verify if the AI answer cites the latest version of the content.
- Ask questions about SOP documents containing complex tables and multi-level headings. Check if the AI answer can accurately extract key information from tables and point the citation source to the page number or paragraph where the table is located.
The values provided are common starting points and should be measured against specific 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.