Citing Sources and Traceability for SMO Regulations

Data Characteristics for this Category

SMO (Site Management Organization) regulatory documents primarily originate from the National Medical Products Administration (NMPA), institutional ethics committees, and internal Standard Operating Procedures (SOPs) developed by SMO companies. These documents are typically in PDF, Word, and Excel formats. Most are unstructured or semi-structured text. Updates usually occur annually or are revised irregularly based on regulations or institutional requirements. Document content covers clinical trial ethical review, informed consent, drug management, adverse event reporting, and data management. Fields include, but are not limited to, regulatory clause numbers, SOP numbers, version numbers, effective dates, revision histories, responsible persons, operating procedures, and sample record forms. Some documents may contain flowcharts or tables, but the content is predominantly text-based descriptions.

Constraints from these Characteristics on "Citing Sources and Traceability"

SMO regulatory documents have a relatively low update frequency. However, each revision may involve core operating procedures or compliance requirements. This necessitates that source citations must be precise down to the revision version. Diverse and heterogeneous document formats, especially scanned PDFs or complex tables, can lead to inaccurate text extraction, affecting RAG recall quality and traceability accuracy. Key metadata such as SOP numbers and version numbers are scattered throughout documents, requiring more refined parsing strategies to ensure citations point to the correct version. Additionally, due to high compliance requirements, answers must provide clear original text citation locations (e.g., page numbers, paragraphs) to support subsequent manual verification and auditing. Answers lacking accurate traceability may be considered invalid in practical applications and could even pose compliance risks.

Configuration Settings

Configuration Item	Recommended Value	Rationale for Recommendation
`Chunk size`	`500–800 characters`	Balances semantic integrity of paragraphs with recall efficiency, preventing overly long paragraphs from diluting key information.
`Chunk Overlap Length`	`50–100 characters`	Ensures contextual continuity at paragraph boundaries, improving cross-paragraph semantic understanding.
`Recall count`	`Top 5–8 entries`	Balances recall breadth with computational cost for complex regulatory inquiries.
`Similarity threshold`	`Calibrate by actual measurement`	Calibrated based on specific corpus and model performance to ensure relevance of recall results.
`Rerank result count`	`Top 3 entries`	Focuses on the most relevant content, reducing the model's burden of processing irrelevant information and improving answer quality.
`citation Information Display`	`Automatic Display`	Mandatory for regulatory Q&A scenarios to ensure compliance.

Three Common Mistakes

Answers displaying "no permission to operate this conversation record" (No permission to operate this chat record) or not showing citation information. This usually results from frontend configuration or backend permission settings, such as the trace_source field not being passed correctly or the frontend component not rendering properly.
Answer content directly outputting \n without proper line breaks. This indicates that the newline characters from the model's output are not correctly parsed as HTML <br> tags or CSS white-space: pre-wrap styles during frontend rendering.
Cited sources pointing to a version that does not match the actually effective version. This can happen if version information in metadata is not correctly extracted or associated during document parsing, leading to outdated or incorrect versions of regulations being indexed and cited.

How to Confirm Correct Configuration

Query different versions of SOPs and verify that the SOP numbers and version numbers cited in the answers match expectations.
Randomly select 10 answers, and for each, verify that the cited original text paragraphs indeed support the answer content and that the cited page numbers or paragraph locations are accurate.
Submit regulatory questions containing ambiguous terms or vague descriptions. Observe the similarity values of the recall results and determine if the top few recalled items are truly relevant.

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.