Citing and Tracing Sources in Neurodegenerative Disease

Data Characteristics

Regulatory and Standard Operating Procedure (SOP) data in the neurodegenerative disease field originates from guidelines, clinical trial protocols, drug development specifications, and related laws and regulations. These are issued by regulatory bodies such as the National Medical Products Administration (NMPA) and the U.S. Food and Drug Administration (FDA). Documents are typically in PDF, Word, or scanned image formats. Update frequency is irregular, but concentrated updates occur when new drug approvals or clinical trial methodology revisions happen. Document structures are complex, containing extensive specialized terminology, acronyms, charts, and cross-references. Common fields and units include dosage units (mg/kg, IU), time units (weeks, months, years), concentration units (nM, μM), and various biomarker metrics. Documents often include reference lists pointing to underlying research data or historical regulations.

Constraints on "Citing and Tracing Sources" from These Characteristics

The complexity of neurodegenerative disease regulatory documents imposes specific requirements on citation and traceability. First, the prevalence of specialized terminology and acronyms demands that the tokenizer accurately recognizes domain-specific vocabulary to prevent citation deviations caused by incorrect tokenization. Second, multi-level cross-references and reference lists mean that traditional paragraph or sentence-based citations may not capture the complete logical chain. Support for tracing citation chains is necessary. Uncertain update frequencies require the knowledge base to have version management capabilities, ensuring that citations refer to the latest effective version of the regulation. Charts and tabular data within documents may lose their contextual structure after text extraction, making it difficult to accurately point to the original chart during citation. This requires considering the preservation of structured information during text preprocessing. Additionally, if documents are scanned images, OCR accuracy directly impacts the completeness and credibility of cited content.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Size)	`800–1200 characters`	Balances contextual completeness with retrieval efficiency, avoiding dilution of core information in overly long chunks.
`Chunk Overlap Length` (Overlap Size)	`150–200 characters`	Ensures continuity of information across chunks, capturing potential cross-chunk citation relationships.
`Recall count` (Recall Count)	`Top 8–12 items`	Given the complexity of regulatory documents, increasing the recall count covers more potentially relevant content.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Ensures recalled content is highly relevant to the query, filtering out low-quality or inaccurate citation candidates.
`Rerank result count` (Reranked Return Count)	`Top 5 items`	Focuses on displaying the most relevant citations, improving user efficiency in obtaining information.
`Document Parsing Timeout`	`300 seconds`	Addresses the parsing needs of large regulatory documents, preventing parsing failures due to oversized documents.

Common Pitfalls

Garbled citation numbers or inaccurate citation links in AI dialogue output often result from incorrect handling of special characters or formatting during text preprocessing, leading to citation anchor shifts.
When asking questions not present in the knowledge base, the AI still provides citations from knowledge base files. This may occur if the Similarity threshold (Similarity Threshold) is set too low, causing irrelevant content to be recalled and mistakenly identified as a citation source.
The absence of cite citation IDs in the JSON returned by the dialogue request interface is often due to improper configuration of the tool calling module in the workflow, or the knowledge base connector not passing citation information as an output field.

Verification Steps

Select a neurodegenerative disease regulatory document containing various structures (text, lists, tables). Conduct multiple Q&A tests, verifying that the cited passages in each answer accurately point to the corresponding location in the original text.
Ask questions specifically targeting specialized terminology and acronyms within the document. Check if the AI can correctly identify and provide citations containing these terms.
Simulate asking general questions not included in the knowledge base. Observe if the AI avoids providing misleading knowledge base citations to validate the effectiveness of the Similarity threshold (Similarity Threshold).
In the FastGPT interface or via API calls, check if the cite field in each dialogue response contains the correct document ID, paragraph ID, and precise cited text. Compare this information with the original document.

Note: The values provided are common starting points. Measure them 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.