Multi-Turn Conversations and Prompts for Hematologic

Data Characteristics

Data for hematologic oncology clinical trial pre-screening comes from multiple sources. Clinical trial protocols are central, typically published as PDF or Word documents. These documents detail trial objectives, inclusion/exclusion criteria, treatment plans, and assessment metrics. Patient electronic medical records (EMRs) are another key source. EMRs include diagnostic reports, genetic test results, imaging reports, laboratory tests (e.g., complete blood counts, bone marrow biopsy reports), and prior treatment history.

Data update frequencies vary. Clinical trial protocols are relatively stable, but amendments are released periodically. Patient EMR data updates in real-time as treatment progresses. Key fields include hematologic oncology-specific diagnostic information such as WHO classification, FISH/gene mutations (e.g., FLT3-ITD, IDH1/2), and chromosomal karyotypes (e.g., Ph+). Units involve common laboratory metrics (e.g., g/dL, cells/µL) and tumor burden assessment (e.g., %).

Constraints on Multi-Turn Conversations and Prompts

The complexity of clinical trial protocols requires multi-turn conversations to deeply understand long texts and continuously reference specific paragraphs. The dynamic nature of patient EMR data, especially updates to genetic mutations and treatment history, demands prompt designs that effectively handle information changes and support version tracking. Hematologic oncology-specific diagnostic fields, such as particular gene mutations, must be explicitly mentioned in prompts to ensure the model accurately matches inclusion/exclusion criteria.

Multi-turn conversations need to support tracing a patient's historical treatment plans, for example, by asking "What was the patient's last treatment plan?" and adjusting subsequent pre-screening logic based on the answer. Additionally, since patient EMR data may contain unstructured descriptions, prompts must guide the model to extract and normalize information. An example is converting "bone marrow blast percentage greater than 20%" into structured data for inclusion/exclusion decisions.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8192` tokens	Addresses the long-text context requirements of clinical trial protocols
`Recall Count`	`10` items	Ensures coverage of multiple data sources, improving matching accuracy
`Similarity Threshold`	`0.75`	Balances recall and precision, filtering irrelevant information
`Rerank Return Count`	`5` items	Focuses on the most relevant inclusion/exclusion criteria and patient features
`Temperature`	`0.1`	Aims for deterministic and consistent results, reducing hallucinations
`Segment Length`	`500` characters	Optimizes long document retrieval efficiency, preventing segment overload

Common Pitfalls

During multi-turn conversations, the model fails to correctly identify updated treatment information in patient EMRs, leading to recommendations for inapplicable clinical trials. This occurs when prompts do not explicitly instruct the model to prioritize data with the latest timestamp, or when insufficient context is provided to distinguish between historical and current states.
The "Get Chat History List" API returns chat records where the correspondence between AI responses and user questions is disorganized. This is due to a lack of unique message_id or parent_message_id for association, making conversation flow tracking difficult.
During a conversation, the model cannot modify a global variable's value as specified by the user, and subsequent conversations still use the old value. This happens when prompt design fails to clearly define variable update triggers and logic, or when there are flaws in the component-to-component variable passing mechanism.

Verification

For a specific hematologic oncology type (e.g., Acute Myeloid Leukemia AML), select multiple anonymous patient EMRs and corresponding clinical trial protocols. Simulate multi-turn conversations to verify if the model can correctly determine inclusion/exclusion criteria. Record the consistency ratio between model judgments and expert judgments.
Check the data returned by the "Get Chat History List" API. Confirm that each user question and AI response is correctly associated via parent_message_id or other identifiers, forming a clear conversation chain.
Simulate scenarios where global variables are modified during multi-turn conversations, such as changing a patient's gene mutation status. Observe whether the model accurately references and uses the updated variable value in subsequent conversations.
Check if the model's understanding and use of hematologic oncology-specific terminology (e.g., Ph+ ALL, CAR-T) in its responses are accurate, ensuring semantic consistency with the medical professional context.

The values provided are common starting points. Measure 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.