Reference and Traceability for Market Access

Data Characteristics

Market access pharmacovigilance data primarily comes from Marketing Authorization Holders (MAH) submissions. These include Risk Management Plans (RMP), Periodic Safety Update Reports (PSUR), Individual Case Safety Reports (ICSR), and drug safety announcements from regulatory bodies. Documents are typically in PDF format, containing extensive unstructured text, tables, and figures. Data update frequencies vary. PSURs usually update semi-annually or annually. RMPs adjust dynamically based on risk assessment results. ICSRs are continuous, real-time submissions. Fields within these documents include drug name, active ingredient, indications, adverse event terms (e.g., MedDRA codes), severity, occurrence date, and reporting source. Units for dosage are commonly milligrams (mg) or grams (g). Frequencies are expressed as once daily (QD) or twice daily (BID). Time units include days, weeks, months, and years.

Constraints on "Reference and Traceability"

The unstructured and complex nature of market access pharmacovigilance documents challenges knowledge base reference and traceability capabilities. Extracting text from PDF tables and figures often loses structural information, impacting subsequent semantic understanding and precise referencing. Diverse document types and update frequencies require systems to handle incremental updates and version management flexibly, ensuring references always point to the latest authoritative versions. The real-time nature of ICSRs means the knowledge base needs to support high-frequency data ingestion and rapid indexing. The presence of specialized terminology, such as MedDRA codes, demands industry-specific tokenization and semantic matching to avoid inaccurate references due to terminology misunderstandings. Given the stringent requirements for this type of data, each reference must explicitly point to a specific page number or paragraph in the original document to meet compliance review needs.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`Chunk size` (Chunk Size)	`500–800 characters`	Balances semantic completeness with recall efficiency, ensuring sufficient context per chunk.
`Overlap Size`	`50–100 characters`	Maintains contextual continuity between chunks, reducing information loss.
`Recall count` (Recall Count)	`10–15 items`	Covers a broader range of potentially relevant information while controlling LLM input length.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Filters out low-relevance results, improves reference quality, and reduces irrelevant information interference.
`maxContext`	`3500–4000 tokens`	Accommodates the detailed descriptions in market access documents, ensuring the LLM can process sufficient context.
`Reference Display`	`Original Document Name + Page Number`	Meets regulatory compliance requirements, providing a precise traceability path.

Common Pitfalls

Symptom: Knowledge base answers lack or have inaccurate page numbers for cited source documents. Reason: Page number information was not correctly extracted or preserved during document parsing, or chunking logic split content across pages.
Symptom: Model-cited content does not align semantically with the original document, or "hallucinations" occur. Reason: The Similarity threshold (Similarity Threshold) is set too low, leading to the recall of numerous low-relevance or ambiguous snippets.
Symptom: When processing a market access query, the model outputs too many reference documents, potentially exceeding the LLM's context limit. Reason: The Recall count (Recall Count) is set too high, and re-ranking mechanisms are not effectively used to filter redundant information.

Verification Steps

Select a typical RMP document, ask questions, and verify that the cited page numbers in the answer accurately point to the corresponding content in the original text.
Query a PSUR document containing complex tables, and verify if the model can correctly understand and cite data from the tables.
Randomly sample multiple Individual Case Safety Report (ICSR) references, confirming that their Reference Display format is consistent and traceable to the original report.
For documents with different update frequencies (e.g., RMP and PSUR), update their content and then query to verify if the knowledge base cites the latest version of the information.

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.