Source Citation and Traceability for CMC Research

Data Characteristics

CMC research regulations and SOP documents are typically in PDF, Word, or internal knowledge base page formats. These documents are detailed, containing extensive technical specifications, experimental methods, quality standards, and compliance requirements. Data update frequency is relatively low, primarily occurring with regulatory updates, technological advancements, or internal process optimizations, usually quarterly or annually. Document structures are rigorous, often including chapters, sub-sections, figures, tables, and appendices. They may also reference external regulatory files or other internal SOPs. Fields and units are highly specialized and standardized, covering specific compound names, batch numbers, analytical method IDs, detection limits, concentration units (e.g., mg/mL, %), temperature units (°C), and time units (hours, days).

Constraints on Source Citation and Traceability

The rigor and specialization of CMC research documents require precise citation of original source locations to ensure answer accuracy and verifiability. The low document update frequency means knowledge base updates must carefully handle version differences to avoid citing outdated information. Complex document structures and specialized fields increase the difficulty of chunking and indexing, requiring the AI to accurately identify and extract key information segments. The abundance of specialized terminology and specific units demands higher semantic understanding from text embedding models to prevent citation errors due to ambiguous terms. Furthermore, the emphasis on regulatory compliance means every answer must be traceable to its original regulatory clause, which is critical for the stability and reliability of the traceability mechanism.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk Size`	`500-800 characters`	Balances information completeness within a single chunk and relevance during recall, preventing noise from overly long chunks.
`Chunk Overlap`	`50 characters`	Ensures contextual continuity at chunk boundaries, improving the success rate of cross-paragraph information retrieval.
`Recall Count`	`Top 5-8`	Covers sufficient relevant information sources while avoiding the introduction of too many irrelevant document segments that could affect inference efficiency.
`Similarity Threshold`	`0.75-0.85`	Improves the precision of recall results for CMC professional terminology and standardized expressions, reducing miscitations.
`Rerank Return Count`	`Top 3`	Further refines the most relevant citation segments after optimization by the reranking model, enhancing answer quality.
`Citation Display Strategy`	`Link at paragraph end`	Ensures each citation can be directly traced to the specific location in the original document, meeting compliance requirements.

Common Pitfalls

Knowledge base answers display "No permission to operate this chat record" at the end of paragraphs, or citation links are unclickable. This indicates the backend service has incorrect file access permissions or the frontend fails to parse citation paths correctly.
Answers are expected to have line breaks but output \n. This occurs when the text rendering component does not correctly identify and convert line breaks, or the model's output format does not match frontend requirements.
Answer content significantly deviates from the cited source, or an irrelevant document segment is cited. This happens when the Similarity Threshold is set too low, leading to the recall of irrelevant documents, or the Rerank Model fails to effectively filter low-quality results.

Verification Steps

Select several representative CMC regulatory questions and submit them to the knowledge base. Check if the cited sources in the answers accurately point to the corresponding paragraphs in the original documents.
Verify that each citation link is clickable and correctly opens or navigates to the precise location in the original document.
Test regulatory documents containing images, tables, or special formatting. Confirm that the AI's citations of their content are accurate.
After a knowledge base update, retest the same questions to confirm that citation sources are not outdated or incorrectly pointing to older document versions.

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.