Knowledge Base Retrieval and Recall for Stem Cell Therapy

Data Characteristics

Stem cell therapy regulations and SOP documents originate primarily from normative documents published by national drug administration and health commissions, industry association guidelines, and internal operating procedures from various medical institutions. Document updates are relatively stable, typically revised after new technological breakthroughs or regulatory policies. Annual or multi-year major revisions are common. Document structures generally use hierarchical, chapter-based layouts, including introductions, definitions, responsibilities, operating procedures, quality control, and risk management modules. Specific fields and units often involve medical terminology, biological indicators (e.g., cell viability percentage, cell count units like cells/mL), time periods (e.g., culture time: 72 hours), and specific approval process codes or batch numbers. File formats are mainly PDF and Word documents, with a significant proportion of scanned images.

Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall

The update frequency of stem cell therapy regulatory documents requires the knowledge base to have efficient version management and incremental update capabilities to ensure retrieval timeliness. The hierarchical document structure demands a chunking strategy that effectively identifies and maintains the semantic integrity of chapters, preventing incorrect truncation at critical operating steps or definitions. Medical terminology and biological indicators within documents require more specialized tokenizers and vector models; general models may struggle to accurately capture semantic relationships. The presence of scanned images necessitates OCR technology for text extraction, which can affect text quality and, consequently, vectorization and retrieval accuracy. Specific approval process codes and batch numbers may require exact or regex matching to supplement fuzzy semantic retrieval.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk Length`	`800–1200 characters`	Balances semantic integrity and recall efficiency. Operating procedures in stem cell therapy SOPs are often long; overly short chunks can fragment context, while overly long chunks may introduce too much noise, affecting recall precision.
`Overlap Length`	`100–150 characters`	Ensures contextual continuity between adjacent chunks, reducing semantic loss due to chunk boundaries, especially for complex biological principles or operational procedure descriptions.
`Parsing Strategy`	`Combine by Title Hierarchy and Fixed Length`	Prioritizes identifying chapter titles (e.g., `H1`, `H2` tags) to maintain the logical structure of regulatory documents, then uses fixed-length splitting for paragraphs without clear titles.
`Recall Count`	`Top 5–8 items`	Accuracy requirements for stem cell therapy regulatory Q&A are very high. Increasing the recall count can improve coverage of relevant information, with subsequent re-ranking models filtering for the most relevant results.
`Similarity Threshold`	Calibrate by measurement	Requires multiple tests with specific embedding models and datasets to ensure highly relevant text is recalled while filtering out noise. An initial setting of `0.75` can be optimized iteratively.
`Rerank Return Count`	`3 items`	After recalling multiple candidate texts, a re-ranking model further refines the selection to the 3 most core and relevant pieces of information, directly supporting the large language model's answer.

Three Common Pitfalls

Retrieval results contain excessive irrelevant information or lack critical details. This often stems from an improper chunking strategy that fails to effectively recognize document structure, leading to semantic fragmentation or excessive noise.
After uploading PDF documents, retrieval quality is significantly lower than expected, or garbled text appears. This usually occurs when PDFs are scanned images without OCR processing, or when OCR quality is poor, resulting in inaccurate text extraction.
After deploying a re-ranking model, the Rerank Return Count for retrieval results is consistently false or does not meet expectations. This typically indicates that the re-ranking model configuration is not correctly activated, or the Rerank Return Count parameter setting does not match the actual model output.

How to Confirm Correct Configuration

Select representative queries for the category and test them in the knowledge base Q&A interface. Verify that the returned recalled text accurately covers the core of the question and includes relevant medical terminology and specific indicators.
Check the knowledge base administration backend. Confirm that the chunk preview, with Chunk Length and Overlap Length parameters set, meets expectations, especially regarding the completeness of critical operating steps.
Through logs or API return results, verify that Recall Count and Rerank Return Count match the configured parameters, and observe if the re-ranked results show a significant improvement in relevance ordering.

Note: The values provided are common starting points. Always measure against your own samples to determine the optimal configuration.

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.