Multi-turn Conversations and Prompts for Hematologic

Data Characteristics in this Category

Hematologic oncology data comes from various sources. These include clinical guidelines, drug inserts, medical literature (such as PubMed, ASCO annual meeting reports), clinical trial data, and patient case reports. Data updates frequently. New drug approvals, clinical trial results, and treatment plan revisions often lead to major updates quarterly or semi-annually. Document structures vary. Drug inserts follow strict regulatory templates, including standardized fields for indications, dosage, and adverse reactions. Clinical guidelines are typically structured text, containing diagnostic criteria and treatment flowcharts. Medical literature primarily uses unstructured or semi-structured text, covering research methods, results, and discussions. Fields and units are highly specialized. Examples include hematological indicators like platelet count (units: 10^9/L), leukocyte differentiation, gene mutation sites (e.g., FLT3-ITD), and drug dosages (e.g., mg/kg).

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

Frequent data updates in hematologic oncology require FastGPT's knowledge base to quickly synchronize the latest information. This prevents multi-turn conversations from citing outdated or incorrect treatment plans. Diverse document structures, especially the large volume of unstructured medical literature, challenge RAG retrieval accuracy and recall. In multi-turn conversations, users might start with a symptom and gradually delve into genetic testing, targeted therapy drug selection, and adverse reaction management. This requires strong context understanding and reasoning capabilities from the model. Specialized fields and units, such as drug interactions and the impact of specific gene mutations on prognosis, demand precise prompt design. Prompts must guide the model to identify and interpret these specialized terms accurately. This ensures professional and accurate answers and avoids information bias due to incorrect unit or dosage interpretation.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`maxContext`	`6` turns	Hematologic oncology product inquiries often involve complex medical histories and treatment plans, requiring a longer context.
`Similarity threshold` (Similarity Threshold)	`0.78`	Ensures recalled medical texts are highly relevant to the query, reducing interference from generalized information.
`Chunk size` (Segment Length)	`800` characters	Balances medical literature information density with model processing capability, preventing semantic loss or excessive length.
`Recall count` (Recall Count)	`10` items	Increases the likelihood of covering relevant medical evidence, addressing complex queries from multiple angles.
`Rerank result count` (Reranked Return Count)	`3` items	Selects the most relevant information to present to the user, improving answer focus.
`Temperature` (`temperature`)	`0.3`	Reduces model divergence, ensuring fact-based answers and preventing speculative medical advice.

Three Common Mistakes

In a multi-turn conversation, the model suddenly provides drug dosages or treatment recommendations that contradict previous discussions. This happens because the knowledge base was not updated in time, and the model recalled outdated drug inserts or clinical guidelines.
When a user asks about the impact of a specific gene mutation on prognosis, the model gives a vague answer. It fails to provide specific clinical evidence or relevant treatment recommendations. This occurs because the prompt lacks clear guidance on the association between "gene mutation" and "prognosis," leading to imprecise knowledge fragment recall.
A token limit exceeded error appears in the OneAPI logs. This is because maxContext is set too high or Chunk size (Segment Length) is too long, causing the number of tokens in a single request to exceed the model's limit.

How to Confirm Proper Configuration

Ask multi-turn questions about the diagnosis, treatment, and prognosis of typical hematologic oncology diseases (e.g., acute myeloid leukemia, multiple myeloma). Check the model's accuracy and consistency, ensuring each conversation correctly cites the latest guidelines.
Consult in detail about specific drug usage, dosage, adverse reactions, and contraindications. Verify that the model's answers, professional terms, dosage units, and values exactly match the latest drug inserts.
Simulate a complex conversation path from symptom description to specific treatment plan selection. Observe the model's performance in context understanding and reasoning. Ensure it can correctly connect information and provide logically coherent advice. Also, check the logs for any FastGPT token authentication failure records.

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.