Data Characteristics
In biopharmaceutical supplier audits, pharmacovigilance data originates from audit reports, supplier-submitted safety data, quality management system documents, contractual agreements, and regulatory compliance records. Document update frequency depends on audit cycles (e.g., annual audits, event-triggered audits) and proactive compliance updates from suppliers. Audit reports typically include executive summaries, findings, observations, corrective and preventive action (CAPA) recommendations, and supporting evidence. Safety data reports cover adverse drug reaction (ADR) and serious adverse event (SAE) reports. Fields include de-identified patient information, drug information, event descriptions, outcomes, reporters, reporting dates, and causality assessments. Drug dosage, frequency, and duration units require precision (e.g., milligrams, days, times). Regulatory compliance records may include batch production records, inspection reports, and change control documents.
Constraints on Knowledge Base Retrieval and Recall
The diverse and variably structured nature of supplier audit data poses multiple challenges for knowledge base retrieval and recall. Audit reports contain unstructured text, while safety reports include semi-structured fields. The knowledge base must handle different data formats. Periodic document updates require support for version management and incremental indexing to ensure timely retrieval. Specialized terminology, drug names, and disease names in safety data demand high precision in tokenization and entity recognition. Traditional keyword-based retrieval may not capture deep semantic relationships. Given data sensitivity, retrieval accuracy and completeness are critical. Insufficient recall or excessive irrelevant information impacts audit decisions. Audits may require cross-document information linking, such as connecting audit findings to specific safety reports or regulatory clauses. The knowledge base must effectively establish logical links between documents during indexing.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 600–800 characters | Balances paragraph integrity in audit reports with retrieval granularity, preventing loss of contextual information. |
Segment Overlap | 100–150 characters | Ensures context continuity across segment boundaries, improving recall for cross-segment queries. |
Recall count (Recall Count) | Top 5–8 items | Balances retrieval efficiency with result comprehensiveness, covering potentially highly relevant document segments. |
Similarity threshold (Similarity Threshold) | Calibrate by testing | Requires testing based on the specific embedding model and data characteristics to ensure high precision and recall. |
Rerank result count (Rerank Return Count) | Top 3 items | Further refines recall results, prioritizing information that best matches the query intent. |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Handles parsing large audit reports or documents with numerous attachments, preventing timeouts. |
Common Pitfalls
- Query results contain many document segments irrelevant to the audit topic. This occurs when the tokenizer fails to effectively recognize specialized biopharmaceutical terminology and abbreviations, leading to generalized matching.
- Knowledge base upload of large supplier audit reports fails with a file parsing error. This is due to the file size exceeding the
UPLOAD_FILE_MAX_SIZElimit or a parsing timeout. - Queries for specific adverse reaction cases fail to recall all relevant safety reports. This may be because the indexing strategy did not fully consider multi-field associations in safety reports, or the
Similarity threshold(Similarity Threshold) is set too high.
Verification
- Execute knowledge base retrievals for typical audit questions. Check if recall results include all expected relevant document segments and verify their accuracy.
- Upload audit files and safety reports in various formats and sizes. Observe if file parsing succeeds and if indexed content is complete.
- Simulate various complex cross-document query scenarios. Evaluate if the knowledge base can effectively link different data types, such as associating an audit finding with its corresponding CAPA or safety report. Adjust the
Similarity threshold(Similarity Threshold) based on business requirements.
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.