Citing Sources and Traceability for Small Molecule

Data Characteristics

Regulatory and SOP documents for small molecule pharmaceuticals originate from drug registration applications, clinical trial protocols, Good Manufacturing Practice (GMP) guidelines, pharmacovigilance procedures, and internal R&D, production, and quality control standards. These documents are typically stored as PDFs, Word files, or Markdown, and are highly structured. They contain extensive specialized terminology, chemical structures, experimental data, and operational procedures. Updates occur quarterly or annually, driven by regulatory changes, new drug development, and manufacturing process optimization. More frequent updates happen for safety-related or significant changes. Documents often include fields like batch number, CAS number, molar mass, and purity percentage, with units precise to multiple decimal places, such as g/mol, %w/w, and ppm.

Constraints Imposed by These Characteristics on "Citing Sources and Traceability"

The rigorous nature of small molecule pharmaceutical regulatory documents demands precise source citation from the Q&A system. Vague or uncertain citations are unacceptable. The specialized terminology and data-intensive content mean traditional keyword-based retrieval methods may miss critical information, requiring deeper semantic understanding. High document update frequency necessitates a knowledge base synchronization mechanism that responds promptly, ensuring the cited regulatory version is current. Complex formats like PDF and Word, which include tables, images, and formulas, require efficient parsing to extract accurate text. For queries involving specific identifiers like batch numbers or CAS numbers, the traceability path must clearly point to the specific paragraph or page in the original document, ensuring answer verifiability.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Size)	`500–800 characters` (characters)	Balances semantic completeness with retrieval efficiency, preventing long chunks from diluting key information.
`Chunk Overlap`	`100–150 characters` (characters)	Ensures contextual continuity and prevents loss of important information due to chunk truncation.
`Recall count` (Retrieval Count)	`Top 5–8 entries` (top 5–8 items)	Covers more potentially relevant paragraphs, improving retrieval accuracy.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Filters out low-relevance content, enhancing answer precision and preventing mis-citations.
`Rerank result count` (Reranked Return Count)	`Top 3 entries` (top 3 items)	Focuses on the most relevant content, reducing the burden on the large language model to process irrelevant information.
`Knowledge Base Version Strategy`	`Latest Version Priority`	Ensures real-time compliance of cited regulations, corresponding to high document update frequency.

Three Common Mistakes

The answer includes input or response details from the knowledge base search, but the user expects only the answer. This usually happens when the large language model (LLM) node in the workflow directly outputs raw tool call results without post-processing or refinement.
When attempting to dynamically specify a knowledge base using a variable in the "Knowledge Base Search" node of the workflow, the Reference Variable dropdown is empty. This indicates the variable is not correctly defined or assigned in the workflow's global or upstream nodes, or the variable type does not match the type expected by the knowledge base selector.
The answer content has minor discrepancies with the original regulatory document, or the cited source points to an outdated version. This may occur if the knowledge base does not synchronize with the latest documents in a timely manner, or if complex structures like tables and formulas are not accurately converted to text during document parsing.

How to Confirm Proper Configuration

For different types of regulatory questions, verify the accuracy of the cited source document name and version number in the answer, and check if the cited paragraph is highly relevant to the answer content.
Select key clauses or data from regulatory documents and ask questions to verify if the system can accurately cite the original text-containing paragraphs or pages.
Simulate a regulatory update scenario by uploading a new document version. Then, verify if the system switches to citing the new document within a reasonable time and confirm the timeliness of citations through questioning.

The values provided are common starting points. Measure against specific samples to determine the most suitable configuration.

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.