Antibody-Drug Conjugates (ADC) Pharmacovigilance: Multi-turn

Data Characteristics for this Category

Antibody-Drug Conjugate (ADC) pharmacovigilance data originates from clinical trial reports, real-world studies, post-market surveillance reports (e.g., FDA Adverse Event Reporting System, FAERS, or EudraVigilance), academic papers, and patient medical records. This data updates frequently, especially during initial drug launch and monitoring periods. Document structures typically follow standardized reporting templates, such as CIOMS I forms or MedWatch 3500 forms. These forms record patient demographics, medication history, adverse event descriptions, event timing, severity, outcome, causality assessment, concomitant medications, and treatment measures. Adverse event descriptions often use medical dictionary terms (e.g., MedDRA) for standardized encoding. Dose units are typically milligrams (mg) or milligrams per kilogram (mg/kg). Time units are days, weeks, or months.

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

The highly structured and standardized nature of ADC pharmacovigilance data requires multi-turn dialogue systems to accurately identify and extract specialized terminology, such as MedDRA codes. This ensures precision in adverse event descriptions. Diverse data sources and high update frequency necessitate an efficient incremental update mechanism for the knowledge base, along with the ability to parse various document formats (PDF, XML, text). Numerical fields like dosage and time within documents require careful prompt design to guide the model in quantitative analysis and comparison. Furthermore, adverse event reports often involve complex causality judgments. The dialogue system must use multi-turn questioning to progressively clarify context and avoid misinterpretations. For ADC-specific complex mechanisms like off-target toxicity and bystander effects, prompts need to guide the model to deeply analyze their association with adverse events, improving risk assessment accuracy.

Configuration Settings

Configuration Item	Recommended Value	Rationale for this Value
`Chunk size` (Segment Length)	`500-800 characters` (characters)	ADC reports often contain lengthy details in adverse event descriptions. This length helps preserve contextual completeness.
`Recall count` (Recall Count)	`Top 8-12 entries` (top 8-12 entries)	Covers multi-dimensional information (patient, drug, event, causality) to ensure comprehensiveness.
`Similarity threshold` (Similarity Threshold)	`0.75-0.85`	Ensures recalled results are highly relevant to the query intent, reducing interference from irrelevant information.
`Rerank result count` (Reranked Return Count)	`Top 5 entries` (top 5 entries)	Focuses on the most relevant key information, reducing model processing burden and improving response speed.
`Max Response Tokens`	`1500-2500`	Ensures the model has sufficient space to generate detailed risk assessments and recommendations, preventing truncated replies.
`Custom Vocabulary`	`MedDRA Terminology`	Accurately identifies and standardizes adverse event descriptions, improving term recognition accuracy.

Three Common Mistakes

Dialogue replies are exceptionally brief, failing to provide a complete analysis. This may be due to a Max Response Tokens setting that is too low, causing the model to truncate before generating critical information.
The system fails to recognize specialized terms for specific adverse events in reports, leading to inaccurate recall results. This occurs when the Custom Vocabulary does not sufficiently cover industry-standard terms like MedDRA.
In multi-turn conversations, the model repeatedly asks for information already provided. This may be due to insufficient maxContext or Historical Message Turns settings, preventing the model from effectively remembering previous dialogue context.

How to Confirm Correct Configuration

Conduct simulated adverse event report queries. Check if the returned risk assessment report covers patient demographics, medication usage, adverse event description, causality judgment, and recommended measures.
Use query statements containing MedDRA terms. Verify if the system accurately identifies and links them to corresponding knowledge base entries. Check the effectiveness of Similarity threshold (Similarity Threshold) and Custom Vocabulary.
Perform multi-turn dialogue tests. Verify if the model can reason and answer based on historical dialogue information after continuous questioning. Confirm that maxContext and Historical Message Turns are configured appropriately.

The values provided are common starting points. Measure performance against specific 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.