Citation and Traceability for Cardiovascular Intervention

Data Characteristics

Cardiovascular intervention regulations and SOP documents originate from national health commissions, medical device regulatory agencies, industry clinical practice guidelines, and internal hospital operating procedures. These documents have a relatively stable update frequency, typically annually or following major policy changes. Documents are often in PDF format, containing numerous tables, diagrams, and flowcharts, such as "Coronary Artery Intervention Guidelines" or "Cardiac Valve Replacement SOP." Text content is highly specialized, involving anatomical terms, surgical instrument models, dosage units (e.g., mg/kg), time units (e.g., min), and detailed descriptions of operating steps. Strong correlations exist between fields.

Constraints on Citation and Traceability

The specialized and structured nature of cardiovascular intervention regulation documents places high demands on citation and traceability. Inaccurate citations of precise instrument models, drug dosages, or operating steps can lead to severe consequences. Therefore, the traceability function must accurately pinpoint specific paragraphs or page numbers in the original text, ensuring information authority. Additionally, tables and flowcharts in documents require special handling during chunking to avoid splitting critical information, which could affect subsequent retrieval and citation. Although the update frequency is not high, each revision may involve core content changes, requiring the knowledge base to quickly identify and update affected knowledge blocks, maintaining the timeliness and accuracy of cited content.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk Length`	`800–1200 characters`	Balances semantic completeness and retrieval efficiency, preventing the splitting of key step descriptions or table rows.
`Recall Count`	`Top 5`	Ensures coverage of multiple relevant regulations or SOPs, improving citation comprehensiveness and preventing omissions.
`Similarity Threshold`	`0.75–0.85`	Balances recall and accuracy, filtering out irrelevant regulations and reducing noise in citations.
`Rerank Return Count`	`Top 3`	Further refines results, ensuring the most relevant core regulations are prioritized for display.
`max_tokens`	`2048`	Reserves sufficient space for the complete presentation of cited content, especially for long paragraphs or table citations.
`Citation Display Mode`	`Paragraph and Page Number`	Must provide precise original document location, facilitating quick verification by engineers against source documents.

Common Pitfalls

Citation responses do not include specific page numbers or paragraphs from the original file: Document chunks are too large, leading to citations only pointing to the entire document, preventing precise localization.
Retrieval results contain numerous irrelevant or outdated regulatory entries: The knowledge base is not regularly cleaned or updated, or the Similarity Threshold is set too low, introducing noise.
Cited content is interrupted or semantically incomplete: Chunk Length is improperly set, splitting table rows or critical operating steps, affecting comprehension.

Verification Steps

Test whether citation responses accurately point to the page numbers and specific paragraphs of the original PDF document for typical cardiovascular intervention operation questions.
Simulate a regulation update scenario. After uploading a new version of the document, verify that related queries' citations have switched to the latest version and check the consistency of the cited content.
Randomly select different types of queries. Check if the number of returned citations matches the Rerank Return Count setting and if the content relevance is high.
Check that when the knowledge base chunk size is set to 5000 tokens, even if max_tokens is limited to 1500, the response still includes citations. However, the cited content will be truncated according to max_tokens. Confirm that the truncated citation remains readable and traceable.

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.