Data Characteristics for this Category
Data for respiratory system products primarily comes from drug inserts, medical device registration certificates, clinical guidelines, academic papers, and internal R&D documents. Update frequencies vary. Drug inserts and registration certificates typically update upon product approval or modification. Clinical guidelines are revised annually or over longer periods, while academic papers are continuously published.
Document structures are as follows:
- Drug inserts and registration certificates are highly standardized. They include fixed fields such as indications, dosage and administration, contraindications, and adverse reactions.
- Clinical guidelines are largely structured text, containing sections on diagnostic criteria, treatment plans, and drug selection.
- Academic papers follow the format of abstract, introduction, methods, results, and discussion.
Common fields and units include:
- Dosage units: milligrams (mg), micrograms (μg), milliliters (mL).
- Time units: hours (h), days (d).
- Disease severity: grades (e.g., mild, moderate, severe) or scoring systems (e.g., GOLD staging, GINA staging).
Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts
The highly standardized and structured nature of respiratory product data allows for more precise extraction and comparison of key information in multi-turn conversations. For example, when asked about drug dosage and administration, the system can directly extract information from fixed fields in the insert.
However, the update cycles and non-standardized text structures of clinical guidelines and academic papers require the knowledge base to have efficient document parsing and semantic understanding capabilities. This addresses variations in versions and expressions.
Qualitative descriptions of disease severity or scoring systems demand more refined prompt design. Prompts must guide users to provide specific numerical values or symptom descriptions, avoiding vague answers.
Furthermore, drug interactions or contraindications for specific patient groups may be distributed across multiple documents. Multi-turn conversations need contextual awareness to synthesize information and provide comprehensive advice, preventing partial responses based on a single source.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 6 | Covers common multi-turn Q&A scenarios, balancing model inference cost and conversational coherence. |
Chunk size | 400-600 characters | Adapts to paragraph lengths in drug inserts and guidelines, ensuring semantic completeness. |
Recall count | Top 5 entries | Ensures relevant key information is retrieved, avoiding interference from excessive irrelevant information. |
Similarity threshold | 0.75 | Balances recall precision and recall rate, reducing irrelevant results. |
Rerank result count | Top 3 entries | Further refines retrieval results, improving the accuracy of the final answer. |
temperature | 0.3-0.5 | Ensures the rigor and accuracy of responses, preventing excessive divergence. |
Three Common Pitfalls
- Workflow interruption: The AI responds, but subsequent processes do not continue. This often occurs when a node in the workflow is misconfigured, causing the returned result format to be unexpected and failing to trigger subsequent nodes correctly.
- Frequent conversation timeouts: This may be due to high latency from integrated external services or model inference time exceeding the configured
PARSE_FILE_TIMEOUT_SECONDSthreshold. - Repetitive plugin prompts: After a user selects a plugin, the prompt message reappears in every conversation. This happens when the plugin prompting logic does not conditionally evaluate based on conversation state or user intent, leading to the same prompt being triggered repeatedly.
How to Verify Correct Configuration
- Conduct multi-turn simulated conversations for different disease classifications (e.g., mild asthma, severe COPD). Check if the system accurately identifies disease severity and provides matching treatment recommendations. Verify the data sources cited in the responses.
- Select specific drug information, such as dosage and administration or contraindications. Intentionally provide vague or incomplete information in multi-turn conversations. Observe if the system can guide the user to provide necessary details through follow-up questions and ultimately provide accurate answers.
- Randomly select 10-20 questions about drug interactions or medication for special populations (e.g., pregnant women, children). Check if the system can synthesize information from multiple documents to provide comprehensive and consistent advice. Verify the consistency of the responses with the latest clinical guidelines.
Note: The values provided are common starting points. Measure them 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.