Multiturn Conversation and Prompts for Solid Tumor Quality

Data Characteristics for This Category

Solid tumor quality documents include clinical trial protocols, investigator brochures, informed consent forms, ethics approval documents, adverse event reports, Good Manufacturing Practice (GMP) documents, and drug registration application materials. Data sources are diverse, encompassing internal pharmaceutical R&D, clinical institutions, and regulatory bodies. Update frequency depends on clinical trial progress, regulatory revisions, and approval status, typically occurring in phases. Document structures are complex, containing specialized terminology, abbreviations, figures, and tables. Fields cover patient characteristics, tumor typing, treatment plans, dosages, efficacy indicators (e.g., ORR, PFS, OS), adverse event grades, biomarkers, and pathology reports. Units often involve dosage (mg, g), time (days, months, years), volume (ml), concentration (ng/mL), and various clinical scoring standards.

Constraints from These Characteristics on Multiturn Conversation and Prompts

The complexity of solid tumor quality documents imposes specific requirements on the accuracy of multiturn conversations and prompt construction. Extensive specialized terminology and abbreviations demand strong semantic understanding from the model to avoid ambiguity and maintain contextual coherence across multiple turns. For example, interpreting the same efficacy indicator may vary across different tumor types; prompts must guide the model to precisely identify the specific context. Phased document updates mean the knowledge base requires regular maintenance to ensure the model responds based on the latest data. Diverse fields and units require prompts to precisely locate information, such as differentiating drug information with various dosage units or filtering by adverse event grades. Complex document structures require the model to effectively extract key information from long texts and flexibly reference different sections during conversation.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`maxContext`	`8000 tokens`	Solid tumor documents are information-dense. A longer context window helps the model maintain a comprehensive understanding of complex cases in multiturn conversations.
`Chunk size` (Chunk Length)	`800–1200 characters`	Balances semantic completeness with vector retrieval efficiency, preventing critical information from being cut off.
`Recall count` (Recall Count)	`Top 8 entries`	Solid tumor queries often involve multiple dimensions. Increasing the recall count improves the coverage of relevant information and reduces omissions.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	While ensuring recall relevance, this moderately relaxed threshold helps capture specialized terms or abbreviations that are related to the query intent but not expressed identically.
`Rerank result count` (Reranked Return Count)	`Top 5 entries`	After reranking, the most relevant core information is prioritized for the model, enhancing the accuracy of multiturn conversations.
`History Turns`	`5–7 Turns`	Discussions about solid tumor diagnosis and treatment plans often require multiple follow-up questions. Maintaining a certain number of historical conversation turns helps the model understand the evolution of user intent and analyze previous discussions in depth.
`Prompt Temperature`	`0.2–0.5`	Given the rigor of quality documents, a lower temperature helps the model generate more precise, factual responses, reducing uncertainty from model creativity.

Three Common Mistakes

Observation: The model fails to accurately link specific patient information or drug dosages mentioned in previous turns of a multiturn conversation. Reason: The History Turns setting is too short, causing the model to lose critical context.
Observation: When a user asks about a specific efficacy indicator, the model's response is generic and does not focus on the interpretation specific to a solid tumor subtype. Reason: The prompt does not explicitly require the model to analyze based on the solid tumor type, or Recall count (Recall Count) or Similarity threshold (Similarity Threshold) are too low, preventing sufficient recall of relevant background information.
Observation: A workflow has multiple AI conversation nodes, but the final output includes responses from all nodes, making it difficult to form a smooth, single conversation. Reason: The workflow design did not effectively manage the output of each AI Conversation node, failing to integrate or filter responses from different nodes and directly exposing all intermediate results.

How to Confirm Proper Configuration

Conduct simulated multiturn conversations covering solid tumor diagnosis, treatment plan comparisons, and adverse event queries. Check if the model consistently understands context and provides highly relevant responses.
Test if the model can provide accurate explanations or contextual information when answering questions involving specific specialized terms or abbreviations. Use this to verify the effectiveness of Similarity threshold (Similarity Threshold) and Chunk size (Chunk Length).
Design queries with complex conditions and cross-references. Verify if the model can extract and integrate information from multiple recalled documents. This helps determine if Recall count (Recall Count) and Rerank result count (Reranked Return Count) are appropriate.

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.