Multiturn Conversations and Prompts for Bispecific Antibody

Data Characteristics

Bispecific antibody registration dossiers draw from diverse and frequently updated data sources. Key data include preclinical study reports (e.g., pharmacodynamics, pharmacokinetics, toxicology), clinical trial data (Phase I/II/III reports, subject case report forms, statistical analysis plans and reports), manufacturing process and quality control documents (CMC, e.g., cell line construction, purification processes, quality standards, stability studies), non-clinical study summaries, and clinical study summaries. These documents are typically in PDF, Word, or Excel formats. They have complex structures, containing extensive specialized terminology, figures, and data tables. Fields and units are highly specific. For example, concentration units might be nM or mg/mL, dosage units might be mg/kg, and time units might be hours, days, or weeks. Data updates primarily follow research and development progress. Clinical trial data updates occur periodically, while CMC documents update after process optimization.

Constraints on Multiturn Conversations and Prompts

The complexity of bispecific antibody dossiers imposes specific requirements on multiturn conversation and prompt design. First, documents contain numerous specialized abbreviations and synonyms. Prompts require robust semantic understanding to prevent information recall failures due to terminology mismatches. Second, data is distributed across various document types and structures. Multiturn conversations must effectively integrate cross-document information. For example, extracting manufacturing batch information from CMC documents and analyzing its correlation with clinical reports. The dense distribution of figures and data tables means that relying solely on text segmentation may not capture complete information. Image recognition or structured table extraction capabilities might be necessary. Finally, uncertain update frequencies require the system to support version management and incremental updates. This ensures conversations are based on the latest, most accurate data, avoiding outdated information.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size`	`800–1200 characters`	Ensures each segment contains sufficient context for complex concepts and long sentences.
`Recall count`	`Top 5-8 entries`	Increases recall coverage due to strong content correlation within documents.
`Similarity threshold`	`0.75–0.85`	Requires high precision for domain-specific terminology to avoid irrelevant recall.
`maxContext`	`4000–8000 tokens`	Manages complex background information and lengthy responses accumulated during multiturn conversations.
`Temperature`	`0.3–0.5`	Ensures accuracy and consistency in responses, reduces hallucinations, suitable for rigorous dossier preparation.
`Rerank result count`	`Top 3 entries`	Further refines recall results, prioritizing the most relevant core information.

Common Pitfalls

When querying stability data for a specific batch, the system returns "No relevant information found." This might occur because batch numbers have inconsistent formats across different documents, preventing precise retrieval.
A user attempts to follow up on a clinical endpoint data point in a multiturn conversation, but the system repeatedly reverts to the initial topic. This happens if maxContext is set too low, causing the model to lose critical context from the conversation history.
After uploading a PDF document with many figures, querying the figures results in an empty or irrelevant text response. This occurs if the file processing pipeline does not include OCR or multimodal recognition for image content, processing only the text layer.

Verification Steps

Conduct multiturn conversation tests for key registration dossier modules (e.g., pharmacokinetic summaries, CMC manufacturing processes). Confirm the system accurately extracts and integrates data from different documents. For example, query Cmax and Tmax values for a specific compound.
Test the system's understanding of specialized terminology abbreviations and synonyms. For example, query using both "双抗" and "双特异性抗体" and observe consistency in recall results.
Simulate questions from an actual dossier review process. Check if system responses comply with professional standards and cite correct document sources and page numbers. This validates the effectiveness of Recall count and Similarity threshold.

The values provided are common starting points. Measure them against your 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.