Multiturn Conversation and Prompts for Cardiovascular

Data Characteristics for This Category

Cardiovascular intervention pharmacovigilance data originates from real-world adverse event reports submitted by medical institutions, clinical trial data for medical device registration, and post-market surveillance reports. This data updates frequently, especially early in a new product's lifecycle, with new reports potentially generated weekly. Document structures typically include structured tables (e.g., CSV, Excel) and unstructured text (e.g., clinical records, follow-up notes). Structured data contains fields such as patient demographics, interventional device model, surgical procedure, adverse event description, and management measures. Unstructured text provides more detailed event specifics, often including medical terminology and abbreviations. Specific fields include Unique Device Identifier (UDI), lot number, and implant site. Units involve device dimensions (millimeters), dosage (milligrams, units), and time periods (days, months, years).

Constraints Imposed by These Characteristics on Multiturn Conversation and Prompts

The diversity and complexity of cardiovascular intervention data require a multiturn conversation system capable of effectively integrating structured and unstructured information. High-frequency data sources mean the knowledge base needs rapid synchronization to ensure the timeliness of conversation results. The extensive medical terminology and abbreviations in unstructured text demand higher semantic understanding from prompts, requiring targeted term mapping and expansion. During a conversation, users may inquire about the incidence of adverse events for a specific device lot. This requires the system to accurately retrieve and aggregate structured data. Interpreting complex adverse event descriptions in clinical records depends on the system's ability to understand long texts and extract key information. Additionally, confirming critical parameters like device model and implant date in multiturn conversations requires prompts to guide the user in providing precise limiting conditions.

Configuration Settings

Configuration Item	Suggested Value	Rationale for This Value
`maxContextLength`	`4096`	Accommodates long text descriptions in unstructured reports, ensuring context completeness.
`Recall count` (Recall Count)	`10-15`	Covers more potentially relevant knowledge, especially when adverse event descriptions are ambiguous.
`Similarity threshold` (Similarity Threshold)	`0.78-0.85`	Balances accuracy and recall, preventing omissions due to medical terminology differences.
`Rerank result count` (Reranked Return Count)	`5`	Prioritizes the most relevant information, reducing user reading burden and focusing on core issues.
`Chunk size` (Segment Length)	`500-800` characters	Optimizes knowledge base retrieval efficiency while ensuring semantic completeness of individual text blocks.
`Max Iterations`	`3`	Limits the depth of multiturn conversations, preventing infinite loops and improving user experience.

Three Common Pitfalls

The conversation unexpectedly terminates midway, with the log showing a KnowledgeBaseQueryFailed error. This occurs because the knowledge base index is not updated in time, leading to queries for non-existent device lot numbers or event codes.
API-driven conversations fail to provide streaming output, causing long waiting times for the frontend. This happens if the API interface is not configured with the stream=true parameter, or if the frontend does not correctly handle Server-Sent Events (SSE) streams.
The output from the AI conversation module cannot be received and processed by subsequent workflow modules. This is due to a mismatch between the output field names of the AI conversation module and the input field names of downstream modules, leading to data transmission interruption.

How to Confirm Proper Configuration

Select typical adverse event reports and conduct multiturn conversation tests to verify if the system accurately identifies device models, adverse event types, and occurrence times.
Simulate user queries about complications for specific device lots. Check if the system can accurately extract and summarize relevant information from structured data and verify the accuracy of the returned results.
Pose questions using clinical cases containing medical abbreviations and complex descriptions. Observe the system's ability to understand unstructured text and determine if key entities are correctly extracted.
Examine conversation logs to confirm if token consumption and response times are within expected ranges, evaluating resource utilization efficiency.

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.