Data Characteristics for This Category
Deviation and CAPA (Corrective and Preventive Action) quality documents originate primarily from pharmaceutical quality management systems, Manufacturing Execution Systems (MES), and Laboratory Information Management Systems (LIMS). These documents typically exist as PDFs, Word files, or scanned images. They are highly structured, containing fixed fields such as event descriptions, investigation results, root cause analyses, corrective actions, preventive actions, and verification results. Update frequency correlates with production batches or quality event occurrences, ranging from daily to weekly or monthly. Once finalized, content modification is infrequent. Documents often include specific terminology, abbreviations, and units, such as "Batch No.", "Deviation Level", "OOS (Out of Specification)", temperature units like "℃", and pressure units like "psi".
Constraints Imposed by These Characteristics on Document Parsing and Chunking
The structured nature of Deviation and CAPA documents requires parsers to identify and extract key information blocks, such as event numbers, occurrence times, and root causes, to ensure precise retrieval. Their update frequency and low modification rate after finalization mean initial parsing requires significant resources for high-quality processing. Subsequent incremental updates primarily involve new documents, with less need for re-parsing existing ones. Industry-specific terminology and units challenge chunking models in understanding contextual semantics. The system must avoid splitting professional terms or separating critical values from their units, which would compromise information integrity. The presence of scanned documents additionally demands high OCR accuracy, especially for tabular data and handwritten annotations.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size | 500–800 characters | Ensures key information (e.g., deviation description, root cause) remains within a single chunk, preventing context loss. |
Chunk Overlap Length | 100 characters | Maintains contextual coherence, especially at logical connections spanning paragraphs. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Accommodates the time required for OCR and structured parsing of large PDFs or complex scanned documents. |
OCR_RECOGNITION_MODE | High-precision mode | Guarantees high accuracy for recognizing specialized terms and tabular data in scanned documents, reducing error rates. |
maxContext | 3000 token | Ensures RAG retrieval can cover complete deviation event descriptions and related actions. |
Three Common Pitfalls
- Uploading large PDF documents results in a "offset out of range" system error. This typically occurs when the file size exceeds the
UPLOAD_FILE_MAX_SIZElimit configured for the server or proxy layer. - After document parsing, some key fields (e.g., "root cause") show empty or incomplete retrieval results. This might be due to the default chunking strategy splitting critical information, or inaccurate OCR recognition of specific fonts or tables within the document.
- Documents uploaded via API show inconsistent chunking results compared to those uploaded through the platform interface. This discrepancy may arise if the
chunk_strategyparameter is not explicitly specified during API calls, leading to the use of a default strategy, while the platform interface might employ a specific, optimized chunking configuration for this category.
How to Confirm Correct Configuration
- Upload representative Deviation and CAPA documents. Check if the content of each chunk in the knowledge base is complete and semantically coherent, without critical information being truncated.
- Perform keyword searches on uploaded documents. Verify that queries containing specialized terms like batch numbers, OOS, and deviation levels accurately recall relevant chunks.
- For scanned documents including tables and handwritten annotations, examine the parsed text content to ensure high OCR recognition accuracy compared to the original, especially for numbers and professional vocabulary.
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.