Data Characteristics
Home healthcare regulations and SOP documents come from various sources. These include medical device regulations from the National Medical Products Administration, industry standards, internal quality management system documents, product manuals, and user guides. Update frequencies vary. Regulations and standards typically revise annually or irregularly. Internal SOPs update monthly or quarterly, depending on product iterations or quality system reviews. Document structures are primarily hierarchical text. Sections often include legal clauses, technical specifications, operating procedures, risk assessments, and troubleshooting guides. Fields and units involve medical device registration numbers, production license numbers, product models, technical parameters (e.g., voltage V, current A, frequency Hz, power W, dimensions mm, weight kg), environmental conditions (e.g., temperature ℃, humidity %RH), and medical units such as dosage mg and treatment duration days.
Constraints on Multiturn Conversation and Prompts
The hierarchical structure and frequent updates of home healthcare regulatory documents require multiturn conversation systems to precisely locate relevant sections and manage differences between new and old versions. Technical parameters and medical units necessitate unit recognition and conversion capabilities to prevent misunderstandings due to inconsistent units. For example, when a user asks about device usage limits within a specific temperature range, the system must extract information like Operating Environment Temperature：5℃～40℃ from the document. Common Q&A pairs and troubleshooting flows in product manuals and user guides provide direct context for multiturn conversations. The strictness of regulatory clauses demands high accuracy in system responses to avoid misleading information. Multiturn conversations must track evolving user intent, such as shifting from "how to sterilize the device" to "criteria for disinfectant selection." This relies on effective management of conversation history.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size | 500–800 characters | Maintains continuity of legal clauses and SOP processes, avoids semantic breaks |
Recall count | 8–12 entries | Covers multiple document sources, ensures relevant regulations and procedures are retrieved |
Similarity threshold | 0.78–0.85 | Balances recall accuracy and rate, reduces irrelevant results |
Rerank result count | 3–5 entries | Focuses on core information, improves response efficiency in multiturn conversations |
maxContext | 3000–4000 token | Supports longer multiturn conversation history, maintains context coherence |
maxTokens | 1024 token | Ensures the model can provide detailed, complete explanations of regulations or SOPs |
Common Pitfalls
- After connecting a local model, the system switches to a cloud model during conversation. This occurs because
LLM_MODEL_NAMEis configured incorrectly orONEAPI_BASE_URLdoes not point to the correct local service address. - The error message "exceeded maximum character limit" appears during conversation. This happens when
maxTokensormaxContextare set too low, preventing the model from processing long texts or multiturn conversation history. - The system cannot accurately answer questions containing specific technical parameters or medical units. This is due to a knowledge base segmentation strategy that does not fully consider the integrity of this information, or the model lacks understanding of unit conversions.
How to Verify Configuration
- Conduct multiturn conversation tests with varying complexities of regulatory and SOP questions. Observe if the system consistently understands user intent and provides accurate answers.
- Select Q&A pairs that include specific values and units. Check if the system output correctly identifies, extracts, and explains this information, for example,
Voltage 220V. - Simulate follow-up questions related to previous conversation topics. Confirm the system provides coherent and meaningful responses based on historical conversation context.
- Check log outputs. Confirm no errors occur during model calls, knowledge base retrieval, and reranking. Verify
tokenconsumption is within expected ranges.
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.