Data Characteristics for This Category
Supplier audit data primarily originates from audit reports, CAPA (Corrective and Preventive Action) plans, supplier qualification documents, production licenses, ISO certifications, and historical audit records. These documents are typically in formats such as PDF, Word, or Excel. Audit reports are usually updated annually or adjusted based on risk levels, while CAPA plans are updated in real-time as remediation progresses. In terms of document structure, audit reports include standardized sections like audit scope, findings, recommendations, and ratings. Qualification documents cover fixed fields such as company basic information, registration number, production scope, and validity period. Key and common fields within this data include dates, batch numbers, specifications, and quality standards (e.g., USP, EP), with units involving measurement units and time units.
Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts
Supplier audit data has a relatively high degree of structure and contains extensive specialized terminology and regulatory requirements. This demands high accuracy in contextual understanding and knowledge recall for multi-turn conversations. Due to the strong correlation between audit reports and CAPA plans, the conversation system needs to integrate information across documents and timelines to answer questions such as "Has the issue identified in the last audit for a certain supplier been rectified?". Managing the validity periods of qualification documents requires the system to identify and track date fields. Furthermore, regulatory compliance is central; prompt design must guide the model to strictly adhere to original information, avoid hallucinations, and ensure the rigor of responses. Accurate identification and explanation of specialized terms, such as "deviation," "batch record," and "GMP," are crucial.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
maxContext | 8000–16000 tokens | Ensures sufficient memory for conversations, covering multi-turn audit questions and relevant document snippets. |
Chunk size (Chunk Size) | 800–1200 characters | Supplier audit reports and CAPA plans often contain detailed descriptions; increasing chunk size helps maintain contextual integrity. |
Recall count (Recall Count) | Top 8–12 entries | Audit questions often involve multiple aspects; increasing recall count improves the probability of retrieving relevant information. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Audit data demands high professionalism; a higher similarity threshold recalls more precise and relevant segments. |
Rerank result count (Reranked Return Count) | 4–6 entries | After initial recall, reranking further optimizes sorting, ensuring the most relevant core information is prioritized. |
temperature | 0.1–0.3 | Audit data Q&A requires high accuracy; a lower temperature helps generate factual, non-creative responses. |
Three Common Mistakes
- When processing questions like "What is the latest audit conclusion for supplier X?", the conversation system fails to accurately identify and prioritize the latest version of the audit report, leading to the citation of outdated information. This occurs because the knowledge base index does not effectively differentiate document versions or lacks a timestamp-based sorting mechanism.
- A user asks, "Please summarize the defect report for batch Y," and the system returns an incomplete defect description or confuses information from different batches. This happens because the prompt does not explicitly include the batch number as a key filtering condition, or document chunking fails to maintain the integrity of batch-specific information.
- During multi-turn follow-up questions, for example, "What is the progress of this supplier's CAPA plan?", the system responds with "Uncaught exception." This might be due to incorrect configuration of the AI conversation component, such as an expired API key or improperly set environment variables in a private deployment, leading to model call failure.
How to Confirm Proper Configuration
- For a specific supplier, simulate asking, "List all major deficiencies found in the most recent audit," and verify if the returned results exactly match the findings in the latest audit report.
- Select an audit report containing a CAPA plan and ask, "What is the corrective action for defect Z, and what is its completion date?" Check if the system can accurately extract and answer with the corresponding action details and date.
- Test multi-turn conversation scenarios, for example, first asking "What is the validity period of supplier A's qualification?" and then following up with "What documents need to be prepared before it expires?" Observe if the system can maintain context and provide relevant regulations or document lists.
The values provided 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.