Knowledge Base Retrieval for Deviation and CAPA Systems

Data Characteristics

Deviation and Corrective and Preventive Action (CAPA) system data originates from pharmaceutical quality management system documents. These include deviation reports, investigation reports, Root Cause Analysis (RCA) documents, CAPA plans, implementation records, and effectiveness verification reports. Documents are typically in PDF, Word, or scanned image formats, with varying degrees of structure. Deviation reports are event-triggered. CAPA plans and execution records update continuously based on the deviation handling process, with cycles ranging from days to months. Document content often includes event descriptions, material batch numbers, equipment serial numbers, operator IDs, timestamps, impact assessment levels, detailed corrective and preventive action steps, and responsible personnel.

Constraints on Knowledge Base Retrieval and Recall

The event-triggered update nature of Deviation and CAPA documents requires the knowledge base to support incremental updates and ensure timely retrieval results. Documents contain extensive specialized terminology, acronyms, and proprietary internal coding rules. This challenges tokenization and entity recognition, potentially leading to low relevance or missed recalls. Complex citation and association relationships exist between different reports. For example, CAPA plans cite specific deviation reports, and root cause analyses link to multiple historical events. The retrieval system must find individual documents and identify and recall related document chains. Timestamps and batch numbers are precise information. The retrieval system needs to extract and match structured or semi-structured information to support precise queries.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Segment Length`	800–1200 characters	Deviation reports and CAPA plans often contain long event descriptions and detailed measures. Longer segments maintain contextual completeness and reduce semantic fragmentation.
`Segment Overlap Length`	100–200 characters	Ensures continuity between segments. This helps the model understand descriptions and causal relationships spanning segments, especially when describing CAPA implementation steps.
`Recall Count`	Top 8	Deviation and CAPA queries often require coverage of multiple related reports. Increasing the recall count improves coverage and reduces missed recalls.
`Similarity Threshold`	Calibrate by measurement; initial value 0.75	Requires tuning based on specific domain corpus and embedding model performance to balance recall and precision.
`Rerank Return Count`	Top 3	After initial recall, a reranking model refines the order of a smaller set of results. This improves the relevance of the final answer and reduces model inference costs.
`PARSE_FILE_TIMEOUT_SECONDS`	300 seconds	Addresses potential time consumption when parsing large PDF or Word documents, preventing parsing interruptions.

Common Pitfalls

Symptom: Retrieval results lack the latest CAPA verification reports. Cause: The knowledge base failed to synchronize recent document updates, or document parsing failed, preventing content from being indexed.
Symptom: A user queries for "deviation report for batch number XX," but no documents with that batch number are returned. Cause: The tokenizer failed to correctly identify the batch number as an independently searchable entity, or the batch number was not processed as a key field during indexing.
Symptom: After deploying the reranking model, retrieval result order does not match expectations, or the reRankScore field is consistently empty. Cause: The reranking function was not correctly enabled in API call parameters, or the rerank_model_id configuration is incorrect, preventing the reranking service from triggering.

Verification

For typical queries, check if recall results include all relevant deviation reports, CAPA plans, and their associated root cause analysis documents. Verify document publication or update timestamps.
Use precise queries containing specific batch numbers, equipment serial numbers, or operator IDs. Verify that the knowledge base accurately recalls documents containing this structured information.
Execute a series of test queries. Observe if the reRankScore field exists and provides reasonable sorting of recall results. Confirm the reranking model is correctly enabled and functioning.
Upload new deviation reports and CAPA records via the API. Check if the knowledge base completes indexing within the specified time. Verify the retrievability of new documents through queries.

The values provided are common starting points. Measure them 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.