Data Characteristics
Data for respiratory system disease policies and SOPs (Standard Operating Procedures) originates from clinical diagnosis and treatment guidelines published by national health commissions, internal hospital regulations, departmental operating specifications, and drug instructions. These documents are typically PDFs, Word files, or scanned images, with varying degrees of structure. National guidelines are usually revised annually or biennially. Internal hospital SOPs are updated periodically based on policy changes or accumulated clinical experience. Documents contain extensive medical terminology, drug names, dosage units (e.g., mg/kg, ml/h), examination indicators (e.g., FEV1, SpO2), and clinical pathways. Some SOPs also include flowcharts or images for illustration.
Constraints on Multi-Turn Conversations and Prompts
The complexity of respiratory system policy documents requires stronger context understanding and entity recognition capabilities for multi-turn conversations. For example, answering questions about drug dosages requires the model to understand the drug name, identify patient-specific parameters like weight and age, and perform calculations based on dosage units in the document. Irregular document updates necessitate an efficient incremental update mechanism for the knowledge base to ensure the timeliness of conversation content. The presence of flowcharts and images in SOPs challenges knowledge base retrieval and multimodal understanding, as pure text retrieval may not fully capture their meaning. Medical jargon and abbreviations require prompt design to standardize terminology and resolve ambiguities, improving answer accuracy.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 500–800 characters | Medical SOP text paragraphs are often long, containing multiple steps or detailed explanations. Longer chunks help maintain semantic integrity. |
Recall count (Retrieval Count) | top 8 | Respiratory disease diagnosis and treatment processes involve multiple steps. Increasing the retrieval count helps cover more comprehensive relevant information. |
Similarity threshold (Similarity Threshold) | 0.78–0.85 | Strict medical policies demand high accuracy. A higher similarity threshold reduces interference from irrelevant or low-relevance content. |
Rerank result count (Reranked Count) | top 5 | In multi-turn conversations, selecting a small number of high-quality context snippets effectively improves the model's focus on key information. |
maxContext | 4096 tokens | This ensures the model can accommodate a sufficiently long conversation history and retrieved content to understand complex medical scenarios. |
temperature | 0.1–0.3 | Policy-related questions require rigorous and factual answers. A lower temperature value helps generate more stable and objective responses. |
Common Pitfalls
- API calls return results missing the
knowledge_source_idfield, preventing tracking of specific cited knowledge sources. This typically occurs because the API call parameters do not explicitly request this field, or source information tracking is not enabled in the knowledge base configuration. - The knowledge base contains flowchart URLs, but the model cannot output or describe image content during conversations. Current RAG models primarily rely on text for retrieval and generation. After embedding image URLs as text, the model lacks the ability to directly parse and describe image content.
- In multi-turn conversations, the model fails to accurately identify specific patient parameters (e.g., weight, age), leading to incorrect drug dosage calculations. This often happens because prompt design does not effectively guide the model to extract key entities from conversation history, or relevant policies in the knowledge base do not explicitly state parameter extraction rules.
Verification Steps
- Conduct multi-turn questioning on typical respiratory disease diagnosis and treatment SOPs (e.g., asthma, COPD). Check the model's accuracy in answering questions about core processes, drug dosages, and precautions. Compare the retrieved content with expected key information for consistency.
- Test questions involving medical terminology and abbreviations of varying complexity. Verify if the model's explanations of terms comply with medical standards. Evaluate its ability to resolve ambiguities in context, such as for
COPD. - Simulate specific patient scenarios (e.g., drug dosage for a child with asthma). Observe if the model can provide logical and evidence-based answers, combining context and document information. Check if the answer cites specific policy sections or clauses.
- Use API calls to confirm the presence of the
knowledge_source_idfield in each conversation's response and verify if it points to the correct knowledge base document ID.
Note: The values provided are common starting points. Measure 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.