Multi-Turn Conversations and Prompts for Medical Imaging

Data Characteristics

Medical imaging device pharmacovigilance data originates from medical institution imaging reports, patient medical records, device usage logs, and adverse event reporting systems (e.g., national drug adverse reaction monitoring centers). This data typically exists as unstructured text (e.g., doctor's descriptions of imaging results, patient chief complaints), semi-structured tables (e.g., adverse event report fields), and structured logs (device operating parameters, error codes).

Data updates occur in real-time for diagnostic reports and medical records as treatment progresses. Adverse event reports have a lag, typically aggregated weekly or monthly. Document structures are complex, containing medical terminology, abbreviations, and device-specific technical parameters. Fields and units include radiation dose (mSv), contrast agent dosage (mL), scan time (s), device models, and serial numbers.

Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts

The highly specialized and diverse nature of medical imaging device data places specific demands on multi-turn conversation accuracy and prompt construction. Unstructured text with medical terminology and abbreviations requires the model to have strong semantic understanding. Prompt design must guide the model to identify and associate specific medical entities.

Device-specific technical parameters and error codes, such as CTDIvol or DICOM error codes, require prompts to precisely target corresponding device manuals or troubleshooting guides in the knowledge base. In multi-turn conversations, users may progressively provide information like device models, scan parameters, and patient symptoms. The model must track context continuously and reason using historical conversation content.

The lag in adverse event reports means knowledge base updates may have a time delay. Prompts need to differentiate real-time information from historical archived data and, if necessary, alert the user to data timeliness.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8`	Ensures the model covers typical multi-turn Q&A scenarios, such as symptom descriptions, device parameter inquiries, and preliminary diagnostic suggestions.
`Chunk size` (Segment Length)	`500 characters` (characters)	Accommodates longer descriptive texts in imaging reports, preventing key information truncation.
`Recall count` (Recall Count)	`10 entries` (items)	Increases the coverage of relevant document snippets retrieved from the knowledge base, addressing the complexity of medical terminology.
`Similarity threshold` (Similarity Threshold)	`0.75`	Balances recall precision and generalization ability, ensuring retrieved documents are highly relevant to the user query.
`Rerank result count` (Reranked Return Count)	`3 entries` (items)	Filters the top 3 most relevant pieces of information for the model, reducing the burden of processing irrelevant information.
`ENABLE_VISION`	`true`	Supports uploading imaging screenshots or device error photos to assist with diagnosis and problem localization.

Common Pitfalls

After a user uploads an imaging screenshot, the model replies, "I cannot provide image content because I am a text-based AI." This occurs because the AI Conversation Node does not have image recognition enabled, or the model itself does not support multimodal input, even if ENABLE_VISION is set to true.
The model quotes many irrelevant historical adverse event reports in its answer, causing the response to deviate from the current device issue. This happens if the Similarity threshold (Similarity Threshold) is set too low, or if the knowledge base contains many general documents with low relevance to the current problem.
A user mentions a device serial number SN:XYZ123 in a multi-turn conversation, but the model cannot find the corresponding device maintenance record. This typically occurs because the prompt does not effectively guide the model to retrieve documents containing specific structured fields (like serial numbers) from the knowledge base.

How to Confirm Correct Configuration

Conduct multi-turn conversation tests for typical imaging device adverse event scenarios. Verify if the model correctly identifies device models, error codes, and provides preliminary handling suggestions.
Upload test cases containing key medical imaging screenshots and device logs. Check if the model recognizes image content and analyzes it in conjunction with text information.
Simulate the user's process of progressively providing information. Observe if the model's understanding of context and its responses are accurate after each conversation, especially when reasoning about numerical fields like radiation dose (mSv) or contrast agent dosage (mL).

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.