Citing Sources and Traceability for Bispecific Antibody

Data Characteristics

Registration documents for bispecific antibody drugs primarily draw data from preclinical study reports, clinical trial reports, manufacturing process validation files, and quality standards with testing methods. These documents are typically in PDF, Word, or Excel formats. They are structurally complex, containing numerous charts, biological sequence information, and statistical data. Data update frequency remains relatively consistent during the R&D to submission phases, primarily updating with new trial batches and interim reports. Beyond standard fields like drug name, batch number, and dosage, these documents also include unique biological indicators such as target affinity (KD value), half-life (t1/2), and immunogenicity evaluation (ADA titer). Units are diverse, including nM, μg/mL, days, and %.

Constraints on Source Citation and Traceability

The complexity of bispecific antibody submission documents imposes high demands on source citation and traceability. Their multimodal data (text, charts, sequences) necessitates robust document parsing capabilities to ensure complete information extraction. For instance, improper parsing of clinical data containing complex tables can lead to incorrect citations of key dosage groups or adverse event rates. Unique biological indicators and their units require the system to accurately identify and present them during citation, preventing unit confusion or misinterpretation of numerical values. Since file update cycles are tied to trial progress, a synchronized knowledge base update mechanism and version management are crucial to ensure that citations always refer to the latest, approved data. Furthermore, internal citation chains between different reports require the system to establish and trace them to verify information completeness and consistency.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Size)	`800–1200 characters`	Individual paragraphs in bispecific antibody submission documents are information-dense. This range helps preserve contextual integrity.
`Recall count` (Recall Count)	`Top 8`	The complexity of submission documents requires more context for accurate judgment and citation.
`Similarity threshold` (Similarity Threshold)	`0.75`	Ensures the precision of recalled content, preventing irrelevant or low-relevance paragraphs from being cited.
`Rerank result count` (Rerank Return Count)	`Top 5`	Further optimizes the order of recall results, placing the most relevant paragraphs at the forefront for citation.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Bispecific antibody submission documents are typically large and structurally complex, requiring longer parsing times.
`maxContext`	`4096 tokens`	Ensures the model has sufficient context space to process lengthy submission document content, preventing truncation of critical information.

Common Pitfalls

The response ends with "No permission to operate this conversation record": This typically occurs when the frontend component attempts to retrieve citation links, but the backend API permission configuration is incorrect or does not return the expected citation ID.
Cited content in the knowledge base answer does not wrap as expected: The model's \n character is not correctly parsed as a newline character during frontend rendering. This might be because the frontend component does not support rich text rendering or does not escape \n characters.
Improper citation limit settings lead to missing key information: If Recall count (Recall Count) or Rerank result count (Rerank Return Count) are set too low, the model may not retrieve enough original text fragments to support the answer, resulting in incomplete or incorrect citations.

Verification Steps

Submit a PDF report containing complex tables and biological indicators. Check if the parsed chunks completely retain the table structure and key numerical values (e.g., KD values).
Ask a question about a specific section in the report (e.g., "Pharmacokinetic Parameters"). Observe if the cited sources in the answer accurately point to the corresponding paragraphs in the original text and if the cited page number is correct.
Deliberately ask a question that requires synthesizing information from multiple sources. Verify if the citation sources in the final answer cover all used original text fragments and if the citation links are clickable and lead to the original text.

Note: The values provided are common starting points and should be measured 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.