Data Characteristics
Supplier audit data in pharmacovigilance primarily originates from audit reports, CAPA (Corrective and Preventive Action) documents, supplier qualification files, quality agreements, adverse event reports, and GMP (Good Manufacturing Practice) compliance certificates. These documents typically exist as PDFs, Word files, or scanned images. Update frequency varies based on the supplier evaluation cycle (usually annual or biennial) and event-driven CAPA processes. Document structures are complex, containing large amounts of unstructured text such as audit finding descriptions, risk assessments, remediation plans, and completion statuses. Fields include audit date, auditor, auditee, defect category, severity, remediation deadline, and responsible person. Units are often dates, text descriptions, or internally defined levels.
Constraints on Multi-Turn Conversations and Prompts
The unstructured nature of supplier audit documents requires FastGPT to use a finer segmentation granularity during vectorization. This ensures critical information, such as specific defect descriptions and corrective actions, is not diluted. Document update frequency dictates the knowledge base synchronization strategy, typically employing scheduled full or incremental updates. This prevents inaccurate conversation results due to outdated information. Professional terminology and abbreviations in audit reports and CAPA documents require prompts to explicitly instruct the model to perform contextual association and explanation. Additionally, in multi-turn conversations, users may trace the remediation status of specific audit findings. This requires the model to extract interrelated information from multiple documents and logically integrate it, for example, identifying issues from an audit report and then confirming resolution from subsequent CAPA documents. Understanding categorical fields like audit finding severity also requires prompts to guide the model in accurately identifying their meaning.
Configuration Recommendations
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size | 300–500 characters | Ensures complete capture of key information like audit findings and corrective actions, reducing information fragmentation risk. |
Recall count | Top 8 entries | Considering the interconnectedness of audit documents, increasing recall to cover more potentially relevant context. |
Similarity threshold | 0.78 | Balances recall and accuracy, filtering out irrelevant background information to focus on core audit content. |
maxContext | 8000 tokens | Accommodates the multi-turn conversation needs of complex audit questions, maintaining a sufficiently long context window. |
System Prompt | Explicitly instructs the model to focus on audit findings, risk levels, remediation plans, and completion statuses, and to explain professional terminology. | Guides model behavior to concentrate on key elements of pharmacovigilance audits and provide necessary explanations. |
Rerank result count | Top 5 entries | Optimizes the quality of knowledge snippets presented to the user, improving answer precision and relevance. |
Common Pitfalls
- The conversation includes "Unable to find the latest audit conclusion for this supplier." This typically indicates the knowledge base has not synchronized the supplier's latest audit report, leading to the model being unable to retrieve the most recent data.
- When answering "Is the CAPA remediation complete?", the model bases its response only on the plan in the audit report, failing to cite subsequent proof of remediation completion. This results in incomplete information.
- When a user asks about the risk level of a specific defect, the model provides a generic answer. It fails to accurately judge based on the levels defined in the document. This occurs because the prompt does not effectively guide the model to identify and apply the classification standards for risk levels within the document.
Verification of Configuration
- Conduct multi-turn conversation tests for typical audit scenarios (e.g., "Query supplier A's GMP compliance status") to verify if the model can accurately cite information from the latest audit report.
- Randomly select audit findings containing complex professional terminology. Test if the model can correctly explain their meaning or provide relevant background knowledge during the conversation.
- Simulate scenarios where a user asks follow-up questions about the remediation status of a specific defect. Check if the model can associate and integrate information from different documents, providing a complete answer that includes the remediation plan and completion status. Verify the accuracy of its information sources.
Note: 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.