Multi-turn Conversations and Prompts for Antibody-Drug

ADC Data Characteristics

Antibody-Drug Conjugate (ADC) data primarily originates from clinical trial reports, patent literature, academic papers, regulatory approval documents (e.g., FDA, EMA), and specialized drug databases. This data is frequently updated, especially during clinical research phases. Document structures typically include target information, conjugation technology, toxin molecules, linkers, drug-antibody ratio (DAR), pharmacokinetic (PK) data, pharmacodynamic (PD) data, and clinical efficacy and safety reports. Fields and units are highly specialized. For example, DAR values are usually unitless integers or decimals (e.g., 2.0, 4.0), half-life is in hours or days, drug concentration is in nanograms/milliliter (ng/mL), and toxin IC50 values are in nanomolar (nM).

Constraints on Multi-turn Conversations and Prompts

The specialized and complex nature of ADC data imposes specific requirements on multi-turn conversation and prompt design. First, high-frequency data updates necessitate rapid knowledge base synchronization to ensure the AI's responses are timely and avoid outdated information. Second, documents contain numerous specialized terms and abbreviations. Prompts must effectively guide the model to understand context, accurately identify entities, and extract relationships. For instance, DAR might refer to drug-antibody ratio, but could have different meanings in other contexts. Prompts must clarify its specific meaning within the ADC domain. Additionally, different fields have widely varying units. The model must ensure unit matching and conversion when generating responses to prevent dimensional errors. During conversations, users may frequently inquire about specific drug PK/PD data or compare efficacy between different ADCs. This requires multi-turn conversations to maintain state, understand implicit connections in subsequent questions, and avoid redundant queries or information loss.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`maxContext`	`6`	Ensures multi-turn conversations cover multiple key stages of ADC drug development, maintaining contextual coherence.
`Chunk size` (Segment Length)	`800–1000 characters`	Accommodates longer specialized descriptive paragraphs in ADC clinical reports and patent documents, reducing semantic fragmentation.
`Recall count` (Recall Count)	`Top 8 entries` (Top 8)	Increases the chance of retrieving relevant ADC data from the knowledge base, especially when multi-dimensional comparisons are involved.
`Similarity threshold` (Similarity Threshold)	`0.78`	Filters out non-core information less relevant to ADC products while ensuring recall rate.
`Rerank result count` (Rerank Return Count)	`Top 5 entries` (Top 5)	Refines search results, prioritizing the most relevant ADC drugs or characteristic information for the user's query.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds` (600 seconds)	Addresses potentially long parsing times for large ADC clinical trial reports or patent documents.

Common Pitfalls

When a user asks about side effects of a specific ADC drug, the AI only returns general toxicity, failing to mention drug-specific adverse reactions. This occurs because detailed toxicology data for that specific ADC is not effectively indexed in the knowledge base, or the prompt does not explicitly request it.
In a multi-turn conversation, a user asks about a specific ADC drug's DAR value, but the AI provides the linker type. This happens because the prompt's disambiguation of the abbreviation DAR is insufficient, leading the model to misunderstand the user's intent.
A workflow execution suddenly aborts mid-way, with the conversation log showing Workflow execution failed: knowledge base search returned no results. This indicates that the knowledge base lacks data for the user's queried ADC product, or the data is not up-to-date.

Verification Steps

Test queries for different ADC drugs to verify if the conversational system accurately provides core information such as targets, toxins, and linkers, and confirm the correctness of specialized terms in the returned information.
Simulate multi-turn conversation scenarios, for example, first asking about an ADC's mechanism of action, then following up with efficacy data in a specific cancer type. Check if the AI maintains contextual coherence and provides accurate answers, comparing with official regulatory agency information.
Randomly select ADC drug data from the knowledge base and pose questions via prompts. Verify if numerical fields in the AI's answers (e.g., IC50 values, half-life) match the original data, and confirm unit correctness.
Check system logs to ensure no PARSE_FILE_TIMEOUT_SECONDS timeouts or knowledge base search returned no results errors occur when processing complex queries or parsing long documents, confirming effective knowledge base utilization.

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.