Data Characteristics in This Category
Hemato-oncology pharmacovigilance data originates from clinical trial reports, real-world studies, adverse event reporting systems (e.g., WHO VigiBase, FDA FAERS), medical literature, and patient case files. This data primarily consists of unstructured text, such as handwritten doctor's notes, patient self-reports, and free-text fields in adverse event report forms. Data updates frequently, especially after new drugs launch, as adverse event reports continuously flow in. Document structures vary, including fields for symptom descriptions, diagnosis results, medication details, and co-morbidities. Information like drug dosage, administration route, adverse event onset time, and duration often appears as numerical values or timestamps. However, descriptive text may also contain vague time or dosage expressions. For example, an adverse event report might state "patient developed a rash approximately one week after taking XX drug" or "twice daily, 10 mg each time."
Constraints Imposed by These Characteristics on Multi-turn Conversations and Prompts
The unstructured nature of hemato-oncology pharmacovigilance data requires stronger semantic understanding from the model in multi-turn conversations to extract key information. High update frequency necessitates that the knowledge base can quickly synchronize new data and update old data to ensure conversation accuracy. Diverse document structures and free-text descriptions demand more refined prompt design. This guides the model to focus on specific fields or information types, for example, instructing the model to distinguish between "drug dosage" and "adverse reaction severity." Furthermore, in multi-turn conversations, patients or doctors may use medical jargon, slang, or non-standard expressions, which challenges the model's robustness. For instance, when a user asks "What is the grade of myelosuppression?", the model must identify "myelosuppression" as a specific adverse reaction from complex text and locate its corresponding severity information.
Configuration Settings
| Configuration Item | Recommended Value | Rationale for This Value |
|---|---|---|
maxContext | 800–1200 characters | Balances long text understanding with computational cost, preventing key information dilution from overly long contexts. |
Recall count | 5–8 entries | Ensures recall of enough relevant knowledge snippets to cover potentially complex queries. |
Similarity threshold | 0.75–0.85 | Filters for highly relevant knowledge under high-precision requirements, reducing interference from irrelevant information. |
Rerank result count | Top 3 entries | Refines the final knowledge presented to the user, focusing on the most core answers. |
Chunk size | 300 characters | Adapts to shorter, information-dense text segments in adverse event reports. |
QUERY_MAX_LENGTH | 200 characters | Limits user input length, preventing semantic dispersion caused by overly long queries. |
Common Mistakes
- The model outputs numerous irrelevant knowledge base reference IDs in conversations, leading to information redundancy. This occurs when the number of recalled entries is set too high and effective re-ranking filtering is absent.
- The model fails to accurately identify key numerical values like dosage or time when answering adverse reaction-related questions. This may be because prompts do not explicitly instruct the model to extract specific entities, or the knowledge base segmentation strategy separates key numerical values from their descriptions.
- Markdown formatting is lost in conversation content when connecting to external platforms, displaying special characters like
#*. This usually happens due to insufficient Markdown rendering support on the external platform or incorrect escaping of special characters during interface transmission.
How to Verify Correct Configuration
- Validate the model's extraction accuracy for key fields such as adverse reaction names, drug dosages, and onset times using a test set, ensuring it meets business requirements.
- Check if the model can correctly understand subsequent questions based on context during multi-turn conversations and provide relevant knowledge base references.
- Simulate user queries and verify that the knowledge base reference IDs returned by the model are highly relevant to the current conversation content and contain no obvious irrelevant information.
Note: The values provided are common starting points. Measure against your own samples to determine optimal settings.
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.