Data Characteristics
Bispecific antibody (BsAb) regulatory and SOP document data primarily originates from internal pharmaceutical company systems. These include quality management, R&D process specifications, clinical trial protocols, manufacturing process documents, and regulatory submission materials. Documents are typically stored in formats like PDF, DOCX, and XML. Content covers the entire lifecycle: target selection, antibody design, cell line construction, fermentation and purification, formulation and filling, preclinical research, clinical trials, pharmaceutical research, non-clinical safety evaluation, quality standards, and stability studies.
Update frequency varies. R&D phase SOPs may revise frequently with project progress. Manufacturing and quality management SOPs usually undergo annual or quarterly review based on regulatory updates or process optimization. Document structures often use chapters, including titles, paragraphs, figures, and appendices. Fields include drug molecular structures, mechanism of action descriptions, dosage units (e.g., mg/kg), batch numbers, expiration dates, and quality control indicators (e.g., Purity %, Endotoxin EU/mg). Units are strict and diverse.
Constraints on "Citing Sources and Tracing"
The complexity and rigor of bispecific antibody regulatory documents demand high capabilities for source citation and tracing. First, multi-format documents require robust file parsing to accurately extract text and table information. Second, frequent updates and revisions make version management and change traceability critical. The RAG system must locate cited content within specific document versions. Third, numerous specialized terms, molecular structures, and measurement units increase semantic understanding difficulty, potentially leading to imprecise recall. For example, a query for "dosage" might require distinguishing clinical from non-clinical dosages. Additionally, complete source citation is crucial. Regulatory Q&A must provide answers and specify the exact SOP and chapter, meeting compliance audit requirements.
Configuration Settings
| Parameter | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Size) | 500–800 characters | Balances semantic completeness and recall efficiency. Avoids noise from overly long chunks and context loss from overly short chunks. |
Chunk Overlap Length (Overlap Size) | 50–100 characters | Ensures contextual continuity, especially in SOP process descriptions, preventing critical information truncation. |
Recall count (Recall Count) | Top 5–8 chunks | Regulatory Q&A requires high accuracy. Increasing recall count helps cover more potentially relevant paragraphs, improving recall rate. |
Similarity threshold (Similarity Threshold) | Calibrate by measurement | Determine through a test set. Balances recall accuracy and relevance, preventing irrelevant documents from being cited. |
Rerank result count (Rerank Count) | 3–5 chunks | Reranks recall results to focus on the most relevant citation snippets, improving final answer quality. |
Citation limit (Citation Limit) | 3 sources | Avoids citing too many sources, maintaining answer conciseness and readability while meeting traceability requirements. |
Common Mistakes
- A "No permission to operate this chat record" message for cited sources usually indicates that the knowledge base has permission control enabled, but the user or system lacks read access to the relevant document.
- No cited sources in the answer may mean the
Citation Source Switchis off in the knowledge base configuration, or the recalled document snippets are insufficient to trigger citation logic. - After exporting a workflow, an "Unable to find plugin referenced in workflow" error when importing into a new environment occurs because the new environment lacks a plugin with the same name as in the original environment, resulting in missing workflow dependencies.
Verification Steps
- Perform multiple rounds of Q&A testing on core regulatory documents. Check answer accuracy and verify that each cited document name and specific paragraph matches the original text.
- Test Q&A functionality with different permission roles. Ensure citation source visibility aligns with expectations, and unauthorized users cannot view sensitive sources.
- Regularly import new or revised SOPs. Test the system's ability to correctly identify and update knowledge base content, and ensure citations point to the latest version.
- Review Q&A logs. Confirm that each Q&A request is accompanied by a valid recall operation, and the recalled document snippets contain expected keywords and context.
The values provided are common starting points. Measure them 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.