Multi-turn Conversations and Prompts for Phase I Clinical

Data Characteristics

Phase I clinical study quality documents include the Protocol, Investigator’s Brochure (IB), Informed Consent Form (ICF), Ethical Review Approval, Investigational Product Management Records, and Adverse Event (AE/SAE) reports. These documents typically originate from the study sponsor, Clinical Research Organization (CRO), or research site. Protocols and IBs may undergo revisions during a study. AE reports and drug management records update in real-time as events occur and daily operations proceed. Documents are primarily in PDF and Word formats, containing extensive structured and semi-structured text. Fields and units involve dosages (e.g., mg/kg), time (e.g., h, day), subject IDs (e.g., SUBJID), and laboratory indicators (e.g., mmol/L, U/L). These exhibit high specialization and strict naming conventions.

Constraints on Multi-turn Conversations and Prompts

The data characteristics of Phase I clinical quality documents impose specific requirements on multi-turn conversation and prompt design. First, the specialized and standardized nature of the documents requires the conversation system to accurately understand professional terminology and abbreviations to avoid misinterpretation. For example, a query about SAE (Serious Adverse Event) requires the system to differentiate its reporting process and determination criteria from general AEs. Second, the dynamic nature of document updates, especially for AE reports and drug management records, demands that the knowledge base rapidly synchronize with the latest information. This ensures multi-turn conversations rely on timely data. If the knowledge base is not updated promptly, subsequent conversations may cite outdated information, leading to inaccurate responses. Third, multi-turn conversations may involve cross-referencing content from different sections of the study protocol. For instance, asking about the dosing frequency for a specific dose group, then following up with the exclusion criteria for that group, requires strong conversational context management to maintain multiple related entities. Additionally, precise identification of specific fields and units, such as querying the unit corresponding to Cmax, requires prompts to guide the model to focus on the relationship between values and units, ensuring accurate information extraction.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Max Segment Length`	`500 characters` (500 characters)	Ensures individual document segments contain sufficient context while avoiding excessive length that reduces model processing efficiency.
`Chunk Overlap Length` (Segment Overlap Length)	`50 characters` (50 characters)	Guarantees contextual continuity and handles concepts spanning across segments.
`Recall count` (Recall Count)	`Top 8 entries` (Top 8)	Phase I clinical documents have strong content interconnections. Increasing recall count enhances coverage for complex queries in multi-turn conversations.
`Similarity threshold` (Similarity Threshold)	`0.78`	Phase I clinical terminology is standardized. A high threshold filters out irrelevant general information.
`maxContext`	`3000 tokens`	Balances multi-turn conversation context length with model processing efficiency, accommodating complex queries.
`Rerank result count` (Rerank Return Count)	`Top 5 entries` (Top 5)	Further refines recall results, improving the model's efficiency in processing relevant information.

Common Pitfalls

Symptom: A user asks about the reporting process for an adverse event, and the AI provides an outdated process version. Reason: The adverse event report document in the knowledge base was not updated in time, leading to multi-turn conversations based on old information.
Symptom: In a multi-turn conversation, the user asks about dosing frequency and exclusion criteria for different dose groups, but the AI cannot link them to provide an accurate answer. Reason: Prompt design failed to effectively guide the model to associate entity information from different document segments in multi-turn conversations, indicating insufficient context management.
Symptom: The SQL query executed by the workflow is successful, but the results are not clearly displayed in the conversation. Reason: When integrating the workflow with the conversation interface, appropriate result_template or display_mode settings were not configured, resulting in a poor display of query results.

Verification Steps

Conduct multi-turn questioning on core professional terms and abbreviations. Verify the AI's accuracy and professionalism, especially its distinction between key concepts like AE and SAE.
Simulate document update scenarios. Upload new versions of study protocols or adverse event reports, then perform relevant queries to confirm the AI can cite the latest information.
Design complex queries involving multiple related entities. For example, "Tell me the Cmax value for SUBJID-001 on day 7, and whether this subject experienced an SAE." Verify the conversation system can correctly associate and extract information.
After a workflow executes an SQL query, check the format and readability of the returned results in the conversation interface. Ensure key fields like dosage, time, and unit are clearly displayed.

Note: The values provided are common starting points. Measure them against specific 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.