Data Characteristics for This Category
Home medical product data primarily comes from product manuals, official website product detail pages, frequently asked questions (FAQs), and user manuals. These documents typically exist as PDFs, HTML, or structured text. Data update frequency is relatively low, mainly occurring during product model iterations, feature upgrades, or regulatory adjustments. Document structures are standardized, including fields such as product name, model, function, target users, contraindications, usage instructions, maintenance, and troubleshooting. Common units include millimeters (mm), grams (g), volts (V), amperes (A), and degrees Celsius (°C). High precision is often required, for example, for blood glucose meter measurement ranges and error values.
Constraints Imposed by These Characteristics on "Multi-turn Conversations and Prompts"
The standardized nature of home medical product data allows for more granular structured segmentation during knowledge base construction, such as slicing by product model or functional module. This improves recall accuracy. Low document update frequency means less pressure on knowledge base synchronization, but historical version management and smooth transitions between old and new knowledge require attention. The high demand for units and precision makes extracting and comparing numerical information critical in multi-turn conversations. Prompt design must guide the model to focus on specific values and avoid vague answers. Furthermore, because products directly relate to user health, accuracy and safety in conversations are core requirements. Prompts must emphasize citing official information and include disclaimers guiding users to consult professional doctors.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk Size | 500–800 characters | Ensures individual knowledge chunks contain sufficient context while avoiding information overload. |
Recall Count | Top 5 | Balances recall breadth with model processing load, ensuring core information is retrieved. |
Similarity Threshold | 0.78–0.85 | Guarantees strong relevance of recalled content, reducing interference from irrelevant information. |
Rerank Return Count | 3 entries | Further optimizes ranking while maintaining relevance, improving user experience. |
maxContext | 3000 Tokens | Accommodates the typical length of home medical product FAQs and supports a certain number of multi-turn conversations. |
response_temperature | 0.3 | Reduces model divergence, ensuring accuracy and rigor in responses. |
Three Common Mistakes
- Phenomenon: The AI response does not cite knowledge base content, or the cited content has low relevance to the user's question. Reason: The
Similarity Thresholdis set too high, preventing relevant knowledge chunks from being recalled; or the knowledge base chunking granularity is too large, with individual chunks containing too much irrelevant information. - Phenomenon: Conversation history is missing or incomplete in MongoDB. Reason: The
record_conversationparameter is not correctly configured totrue, or the storage service connection is abnormal, preventing conversation data from being persisted. - Phenomenon: The AI conversation fails to ask follow-up questions or perform in-depth analysis based on user inquiries. Reason: Prompt design does not effectively guide the model for multi-turn reasoning, for example, lacking instructions like "if the user mentions A, then further inquire about B."
How to Verify Configuration
- Test the AI's accuracy with multiple sets of test questions involving numerical comparisons, function inquiries, and troubleshooting scenarios. Check if the AI cites the correct knowledge base content.
- Use the "Conversation History" feature in the FastGPT backend. Randomly sample conversation history to confirm that conversation turns, user input, AI output, and cited knowledge chunks are all completely recorded.
- In the workflow, simulate multi-turn conversation scenarios for complex problems. Check if the AI can recall and integrate information from the knowledge base based on the previous response or user follow-up questions, forming logically coherent answers.
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.