Multi-Turn Conversations and Prompts for CAR-T Cell Therapy

Data Characteristics

CAR-T cell therapy product data originates from clinical trial reports, drug labels, academic papers, regulatory approval documents, and real-world post-market studies. This data updates infrequently, typically every few months to several years, coinciding with clinical trial progress, new drug approvals, or major research findings. Document structures are complex, containing extensive specialized terminology, abbreviations, and biological pathway descriptions. Common fields include target antigens, cell sources, manufacturing processes, indications, clinical efficacy metrics (e.g., complete response rate CR, objective response rate ORR), adverse event rates AE, dosage, administration protocols, and pharmacokinetic parameters. Units include %, mg/kg, cells/kg, days, months, and years, often accompanied by confidence intervals or statistical significance markers.

Constraints on Multi-Turn Conversations and Prompts

The complexity and specialized nature of CAR-T data impose specific requirements on multi-turn conversation and prompt design. First, the low update frequency necessitates authoritative and timely data sources for knowledge base construction. Prompts must guide the model to prioritize the latest approved information. Second, complex document structures and dense terminology require prompts to include term explanations or contextual linking instructions. This ensures the model accurately understands user intent and extracts key information from unstructured text. For example, if a user mentions "leukemia," the model needs to confirm the specific disease type (e.g., ALL or CLL) through multi-turn dialogue and combine it with target information to provide precise CAR-T product recommendations. Additionally, the quantitative nature of efficacy and safety data dictates that prompts should encourage the model to explicitly cite specific values and units in its answers, avoiding vague statements. For instance, when explaining CR rates, the model should provide percentages and the corresponding study population.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8000` characters	Ensures coverage of key information in lengthy clinical trial reports, reducing information loss due to truncation.
`Chunk size` (Segment Length)	`500–700` characters	Balances semantic completeness of text with recall efficiency, preventing single segments from becoming too long and diluting core information.
`Recall count` (Recall Count)	`8–12` items	Considers the specialized and interconnected nature of CAR-T information, requiring more context for comprehensive judgment.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Improves recall precision, filtering out document snippets with low relevance to highly specialized CAR-T product questions.
`Rerank result count` (Reranked Return Count)	`4–6` items	Further refines the most relevant content from high-similarity recalls, enhancing the quality of the final answer.
`systemPrompt` Instruction	Explicitly requests citation of clinical data	Forces the model to mention data sources, key metric values, and units, such as `CR %`, in its answers.

Common Pitfalls

The conversation includes a large amount of irrelevant or outdated CAR-T product information. This occurs when the knowledge base synchronization mechanism does not effectively link to the latest regulatory approval data, leading the model to cite withdrawn or unapproved products.
When asked about adverse events, the model returns overly general symptom descriptions without mentioning specific incidence rates or grading. This happens when prompts do not explicitly require the model to extract CTCAE grading or specific percentage data for adverse events.
The user modifies the disease type in a multi-turn conversation, but the model continues to recommend products based on the previous turn's disease information. This indicates that the conversation state management mechanism fails to effectively update user intent, leading to context drift.

Validation Steps

Query for different indications (e.g., DLBCL, MM). Check if the model accurately recommends CAR-T products for corresponding targets (e.g., CD19, BCMA) and cites the latest clinical data.
Simulate user inquiries about side effects of specific CAR-T products. Verify that the model's response includes incidence rate ranges or CTCAE grade descriptions for common adverse events.
Conduct multi-turn follow-up questions. For example, first ask about the efficacy of product A, then ask about its differences compared to product B. Observe if the model can accurately compare key metrics of both while maintaining conversational coherence.

The values provided are common starting points. Measure them against your 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.