Data Characteristics for This Category
Hematology-oncology regulations and Standard Operating Procedure (SOP) documents primarily originate from guidelines published by national health commissions and drug administrations, as well as internal rules from medical institutions. These documents update relatively stably, typically every few months to several years. Major revisions occur when new drug approvals, treatment plan updates, or management regulation adjustments are involved. The documents are mostly in PDF and Word formats. Content is rigorous, containing extensive medical terminology, abbreviations, flowcharts, tables, and dosage units. For example, chemotherapy protocols explicitly list drug names, administration routes, dosages (e.g., mg/kg, mg/m²), cycles, and precautions. Diagnostic criteria reference international disease classification codes (e.g., ICD-10) and detail pathological features and laboratory indicators (e.g., PLT, Hb values).
Constraints Imposed by These Characteristics on "Multi-turn Conversations and Prompts"
The rigor and specialized nature of hematology-oncology regulatory documents demand high accuracy in multi-turn conversations. Medical abbreviations and specialized terms in the documents require the model to accurately identify and understand context to avoid misinterpretation due to ambiguity. For example, ALL could mean Acute Lymphoblastic Leukemia or Allergic Reaction; context is crucial for disambiguation. The presence of tables and flowcharts necessitates efficient information extraction capabilities to convert structured data into understandable text. The precision of dosage units (e.g., distinguishing mg/kg from mg/m²) requires the model to accurately restate or calculate, without arbitrary conversion. Furthermore, the cyclical nature of regulation updates requires the knowledge base to quickly synchronize with the latest versions, ensuring the timeliness and compliance of conversation content. Multi-turn conversations may involve cross-referencing multiple regulatory clauses, requiring prompt design to guide the model toward comprehensive judgment.
Configuration Settings
| Configuration Item | Recommended Value | Rationale for Recommendation |
|---|---|---|
maxContext | 8192 tokens | Covers complex medical context while balancing model processing capacity. |
Chunk size (Segment Length) | 500–800 characters | Adapts to regulatory document paragraph length, reducing information fragmentation. |
Recall count (Recall Count) | 8–12 entries | Increases the probability of recalling relevant regulatory clauses, ensuring coverage. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Filters irrelevant content while avoiding omission of critical regulatory details. |
Rerank result count (Rerank Return Count) | 5 entries | Selects the most relevant clauses, improving multi-turn conversation coherence. |
temperature | 0.3–0.5 | Maintains the rigor and objectivity of responses, reducing model hallucination. |
Three Common Mistakes
- Frequent "No relevant information found" responses or overly generic replies during conversations indicate an unreasonable knowledge base segmentation strategy, leading to over-splitting of regulatory clauses or loss of key information.
- The model suddenly "forgets" previous turns in a multi-turn conversation, failing to link to earlier questions. This may be due to the
maxContextparameter being set too low, truncating historical conversation information. - When a user asks about specific drug dosages, the model returns incorrect units or vague descriptions. This happens if dosage fields within tables were not structurally extracted during document preprocessing, or if prompts did not explicitly demand unit precision.
How to Confirm Proper Configuration
- Conduct multi-turn tests with complex questions covering multiple regulatory clauses. Observe if the model accurately cites and integrates information.
- For questions involving medical abbreviations and specialized terms, check if the model's responses correctly explain their meanings and align with the original regulatory text.
- After a regulation update, upload the new version of the document and test related questions. Confirm that the model prioritizes the latest information.
- Design questions that include numerical units for dosage, cycles, etc. Verify that the model's responses provide precise values and units.
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.