Source Citation and Traceability for Market Access

Data Characteristics

Market access regulation documents typically originate from official websites of national drug regulatory agencies (e.g., FDA, EMA, NMPA), industry association guidelines, and legal databases. Update frequency is relatively consistent, aligning with regulatory revision cycles or annual report releases, such as quarterly or annually. Documents are primarily in PDF, DOCX, or HTML formats. Content includes legal provisions, implementation rules, application requirements, approval processes, and technical guidance principles. Key fields include regulation name, issuing authority, effective date, scope, specific article numbers and content, attachment lists, and historical revision records. In specific cases, professional units like product classification codes (e.g., ATC codes), clinical trial phases, and approval timelines are also involved.

Constraints on Source Citation and Traceability

The regulatory nature and rigor of market access documents require precise source citation down to specific articles or paragraphs to support compliance arguments. The periodic updates of these documents mean the knowledge base must regularly synchronize with the latest versions to ensure citation timeliness. Complex document formats like PDF and DOCX present challenges for content parsing and segmentation, especially when handling nested tables, image captions, and cross-references. This can lead to inaccurate segmentation and subsequently affect recall quality. Additionally, the existence of multilingual regulations adds requirements for language matching and terminology consistency in cited content. Fields such as "effective date" and "revision records" are crucial for traceability and must be clearly displayed in citations to avoid referencing outdated or repealed provisions.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters`	Balances the completeness of legal provisions with the precision of recall.
`Recall count` (Recall Count)	`Top 8`	Ensures coverage of relevant provisions from multiple angles while avoiding interference from irrelevant information.
`Similarity threshold` (Similarity Threshold)	`0.75`	Filters out irrelevant recall results, improving citation accuracy.
`Rerank result count` (Rerank Return Count)	`Top 3`	Optimizes sorting results, prioritizing the most relevant provisions.
`maxContext`	`4000 tokens`	Accommodates the lengthy nature of regulatory texts, providing sufficient context for large models.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Handles time-consuming parsing of large regulatory files, preventing processing failures due to timeouts.

Common Mistakes

Chat responses containing "knowledge base search" or raw input/output: This usually occurs when debug parameters like SHOW_QUERY_DETAIL or WORKFLOW_DEBUG_MODE are not disabled.
Cited content displaying "No relevant information found" or missing citations: This may be due to document parsing failure, resulting in no valid segments in the knowledge base, or a Similarity threshold (Similarity Threshold) set too high.
Citing repealed or outdated regulatory provisions: This often happens when the knowledge base is not updated promptly, the knowledge source is out of sync with the latest regulations, or the effective date field is not effectively utilized during vectorization.

Validation

For typical market access questions, verify that the regulation names and article numbers cited in responses exactly match the original document content.
Check the version and effective date of the cited regulations in responses, ensuring they are the latest or user-specified versions, and cross-reference with official release information.
Simulate queries that intentionally include keywords from repealed or revised regulatory provisions. Observe if the system can identify and avoid citing them, or if it indicates they are no longer in effect.
Review backend logs or the debugging interface to confirm that parameters like Chunk size (Segment Length) and Recall count (Recall Count) are functioning as expected in actual queries.

The values given are common starting points and should be measured 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.