Multi-Turn Conversations and Prompts for Solid Tumor

Data Characteristics for This Category

Data for solid tumor-related regulations and Standard Operating Procedures (SOPs) primarily originates from regulatory bodies like the National Medical Products Administration (NMPA), clinical trial institutions, and hospital ethics committees, as well as internal quality management system documents from pharmaceutical companies. These documents update at a relatively stable frequency, typically through annual revisions or ad-hoc updates based on policy changes. Document structures are often hierarchical PDF or Word formats, containing numerous definitions, clauses, flowcharts, and appendices. Fields frequently involve specialized terminology such as disease codes (ICD-O-3), drug batch numbers, clinical stages, and adverse event grades (CTCAE v5.0). Units include dosage (mg/kg), time (hours, days), and percentages (%). Documents are generally lengthy, with single files often exceeding a hundred pages.

Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts

The hierarchical structure and specialized terminology of solid tumor regulatory documents require multi-turn conversation systems to possess context awareness and precise terminology parsing capabilities when understanding user queries. For example, when a user asks, "What is the adverse reaction management process for a specific drug in Phase III clinical trials?", the system must accurately identify "Phase III clinical trials" as stage information and link it to the specific drug's adverse reaction management SOP. The periodic nature of document updates necessitates regular synchronization of the knowledge base with the latest versions to ensure the timeliness of answers. Lengthy documents challenge chunking strategies; overly short chunks may lose context, while overly long ones increase retrieval noise. Furthermore, the standardization of fields and units requires prompt design to effectively guide the model in extracting and presenting information with specific values and units, such as "What is the recommended dosage in mg/kg for this regimen?".

Configuration Guidelines

Configuration Item	Recommended Value	Rationale for Recommendation
`chunkLength`	`800–1200 characters`	Balances contextual completeness with retrieval efficiency, avoiding overly long or short individual chunks.
`recallCount`	`top 5–8 chunks`	Covers highly relevant document segments while reducing interference from irrelevant information.
`similarityThreshold`	`0.75–0.85`	Filters out low-relevance results while ensuring the accuracy of recalled information.
`maxContext`	`3500–4000 tokens`	Ensures the model can process longer historical conversations and retrieved document content.
`rerankCount`	`3 chunks`	Further refines recalled results to improve the relevance of the final answer.
`historyLength`	`4–6 turns`	Maintains coherence in multi-turn conversations, preventing the loss of crucial early information.

Three Common Mistakes

Symptom: After a user query, the system returns "No relevant information found" or provides inaccurate general answers. Reason: The similarityThreshold is set too high, preventing relevant document segments from being recalled, or the document chunking strategy is unreasonable, leading to key information being fragmented.
Symptom: In multi-turn conversations, the model fails to understand the connection between the user's subsequent questions and previous context, resulting in inaccurate answers. Reason: The historyLength is set too low, causing the model to lose historical conversation context when generating responses.
Symptom: After uploading a document, the system displays "File parsing failed, status code 500". Reason: The uploaded PDF or Word document has a complex format, containing numerous images or special fonts, leading to PARSE_FILE_TIMEOUT_SECONDS timeout or the parser's inability to correctly extract text content.

How to Verify Configuration

Select several representative solid tumor regulatory questions, including single-turn and multi-turn follow-ups. Check the accuracy and coherence of the system's answers and compare them with the original documents.
Observe the recallCount and similarity scores for each retrieval in the conversation logs. Ensure that the recalled results cover the key information of the user's question and that the similarity is within a reasonable range.
Test the system's ability to upload and parse solid tumor regulatory documents of various formats and lengths. Check for parsing failures or missing content, and confirm that the chunkLength is suitable for these documents.
For core regulatory clauses, design questions containing specialized terminology and abbreviations. Verify whether the model can correctly understand and cite specific field values and units from the documents.

Note: The values provided are common starting points. It is recommended to 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.