Data Characteristics for This Category
Deviation and Corrective and Preventive Action (CAPA) system documents are typically structured text, such as PDFs, Word documents, or entries in an internal knowledge management system. Data sources primarily include reports, investigation results, analysis reports, and subsequent action plans and execution records entered by quality management or production departments. Update frequencies vary. New deviation events trigger CAPA processes, leading to the creation and revision of relevant documents. Older CAPA actions may be adjusted due to effectiveness evaluations or regulatory updates. Document structures commonly include fields like deviation description, root cause analysis, corrective actions, preventive actions, responsible person, completion deadline, and verification results. Content involves internal identifiers such as batch numbers, equipment codes, and operating procedure numbers, along with specific timestamps and qualitative/quantitative descriptions.
Constraints Imposed by These Characteristics on "Citation and Traceability"
The structured nature of deviation and CAPA documents requires the knowledge base to identify and retain contextual associations of key fields during chunking, ensuring the integrity of cited snippets. The uncertain update frequency necessitates incremental training and version management to ensure retrieval results are current. Documents contain numerous internal identifiers and specialized terms, demanding higher precision in semantic retrieval. The model needs to understand these domain-specific vocabularies. The presence of timestamps and responsible person fields provides clear traceability clues, requiring citation results to clearly point to specific paragraphs in the original document and related metadata. Furthermore, due to the seriousness of CAPA actions, strict accuracy is required for citation sources. Any vague or incorrect citation can lead to serious production or compliance issues.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 500–800 characters | Ensures complete semantic units such as CAPA descriptions, root cause analyses, or action details are contained within a single chunk, preventing key information fragmentation. |
Recall count (Recall Count) | 5–8 items | Balances retrieval efficiency and coverage. Recalls multiple potentially relevant CAPA entries while maintaining relevance. |
Similarity threshold (Similarity Threshold) | 0.7–0.8 | Increases matching precision. Filters out document snippets with low relevance to the query intent, reducing the risk of incorrect citations. |
Rerank result count (Reranked Return Count) | 3 items | After selection by the reranking model, provides the most core and relevant CAPA actions or deviation reports, reducing the burden on the main model. |
maxContext | 3000–4000 tokens | Ensures the large language model receives sufficient context to understand CAPA details and logical flow for accurate summarization and Q&A. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Handles large CAPA summary reports or documents containing complex charts, preventing parsing timeouts. |
Three Common Mistakes
- Citing outdated or revised CAPA document snippets in retrieval results. This typically occurs due to a knowledge base not performing timely incremental updates or having an inadequate version control strategy.
- The AI response cites CAPA actions that appear relevant but do not align with the query intent. This may happen if the
Similarity threshold(Similarity Threshold) is set too low, leading to the recall of many broadly related contents. - The AI cannot provide the responsible person or completion deadline for specific CAPA actions, offering only general descriptions. This stems from a failure to effectively retain the contextual association of key fields during document chunking, or inaccurate field recognition during original document parsing.
How to Verify Configuration
- Submit queries for typical deviation scenarios. Verify that the CAPA documents cited in the AI response are the latest version and that the content is complete and accurate.
- Query for a specific CAPA number. Check if the AI response accurately provides all key information for that CAPA, such as root cause, actions, responsible person, and completion date.
- Simulate misleading queries. Observe if the AI can effectively filter out irrelevant CAPA content, citing only highly matched document snippets, and providing clear citation source links.
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.