Multi-Turn Conversations and Prompts for Hematology-Oncology

Data Characteristics for This Category

Quality documents in the hematology-oncology field originate from pharmaceutical companies' R&D reports, clinical trial protocols, manufacturing process specifications, quality standards, batch production records, inspection reports, and regulatory guidelines. Document update frequency depends on drug development progress, clinical trial results, manufacturing process changes, and regulatory policy adjustments. Updates are typically periodic, such as quarterly or annually. Document structures often include extensive structured data (e.g., test indicators, numerical ranges, batch information) and semi-structured text (e.g., descriptions of test methods, deviation handling records). Common fields include batch number, production date, expiration date, test item, result, unit (e.g., ng/mL, IU/mg, %), and judgment (pass/fail).

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

The periodic update nature of hematology-oncology quality documents requires full index rebuilding or incremental updates after a knowledge base update. This ensures that information referenced in multi-turn conversations is current. The large amount of structured data and specialized units in these documents means prompts must precisely guide the model to recognize and understand the association between values and units. This prevents errors caused by unit confusion. For example, when querying drug concentration, if the model fails to differentiate between μg/mL and mg/mL, it might provide misleading responses. The presence of semi-structured text demands higher capabilities for information extraction and summarization in multi-turn conversations. Prompts need to guide the model to locate key information within complex descriptions and support user follow-up questions on specific details. Furthermore, the specialized nature of the document content requires prompts to effectively constrain the model's responses to specific domain vocabulary and concepts.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`8` turns	Hematology-oncology quality issues often require several follow-up questions to clarify specific details. 8 turns cover most complex scenarios.
`Chunk size` (Segment Length)	`500–700 characters`	Balances structured data and semi-structured descriptions. Prevents overly long segments from diluting key information and overly short segments from losing context.
`Recall count` (Recall Count)	`Top 8`	Ensures coverage of multiple highly relevant document segments in multi-turn conversations, improving information comprehensiveness.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Domain terminology requires high precision. A threshold that is too low may introduce irrelevant content; a threshold that is too high may miss relevant information.
`Rerank result count` (Reranked Return Count)	`Top 5`	Further refines the most relevant segments after initial recall, enhancing multi-turn conversation quality.
`Model Temperature`	`0.3–0.5`	Ensures factual accuracy in responses for rigorous quality document Q&A, reducing generative hallucinations.

Three Common Pitfalls

Unexpected conversation flow jumps: When user input contains ambiguous or polysemous terms, the Question Classification module might route the conversation to an incorrect sub-process. This occurs because the classification prompt lacks sufficient distinction for domain-specific synonyms or near-synonyms.
Redundant history records: When initiating a conversation via API, if the stream parameter or history array is not set correctly, two conversation history records might be generated after a single user request, leading to confused context management.
Model output lacking key numerical values: When querying specific test results, the model's response fails to include specific values or units. This happens because the prompt does not explicitly require the model to extract and return key fields, or the parsing of that field in the knowledge base is not accurate enough.

How to Confirm Proper Configuration

Conduct multi-turn conversation tests. Verify whether the model can accurately maintain context and provide correct answers when asked follow-up questions about key information such as batch numbers, test items, and results.
Examine quality issues of varying complexity. Ensure the Question Classification module accurately directs user intent to the predefined processes. Observe whether the classification results in the logs align with expectations.
Randomly select key numerical values and units from documents. Construct questions and check whether these details in the model's response are complete, accurate, and consistent with the original documents.
Simulate API calls. Observe the generation of conversation history records to confirm that each interaction produces only one history record and that the context parameter is passed correctly.

The values provided are common starting points and should be measured against the reader's 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.