Knowledge Base Retrieval and Recall for Stem Cell Therapy

Data Characteristics

Data for stem cell therapy pharmacovigilance originates from clinical trial reports, real-world evidence (RWE) data, case reports, academic literature, and regulatory adverse event databases. This data updates frequently; adverse event reports are often submitted in real-time or weekly during clinical trials. Document structures vary, including structured Case Report Forms (CRFs), semi-structured medical records, and unstructured free-text descriptions. Fields include patient demographics, stem cell product information (e.g., source, dose, batch), treatment regimens, adverse event occurrence time, type, severity, outcome, and causality assessment. Units involve dosage (e.g., cells/kg), time (e.g., hours, days), and severity grades (e.g., CTCAE grades).

Constraints on Knowledge Base Retrieval and Recall

The unique nature of stem cell therapy imposes multiple constraints on knowledge base retrieval and recall. First, the high heterogeneity of stem cell products (source, differentiation stage, administration method) leads to complex and diverse descriptions of related adverse reactions. The knowledge base must identify semantic connections across different product contexts. Second, the real-time nature of adverse event reporting requires rapid update and indexing capabilities. Third, the presence of extensive unstructured free text means simple keyword matching is insufficient for accurate information recall, necessitating advanced semantic understanding. Additionally, case reports often contain numerous medical acronyms and synonyms, posing challenges for terminology standardization and entity recognition. Adverse event severity and causality assessments involve multi-dimensional information; recall must integrate this context to avoid misinterpretation.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Size)	`500-800 characters`	Balances contextual completeness and retrieval efficiency, accommodating typical medical text paragraph lengths.
`Chunk Overlap Length` (Chunk Overlap)	`50-100 characters`	Ensures contextual continuity at chunk boundaries, preventing critical information from being split.
`Similarity threshold` (Similarity Threshold)	Calibrate based on actual measurements	The complexity of adverse reaction descriptions in stem cell therapy requires adjustment based on real data, typically between `0.7-0.85`.
`Recall count` (Recall Count)	`10-20 items`	Provides sufficient candidate results for subsequent re-ranking and LLM synthesis, while considering retrieval performance.
`Rerank result count` (Re-ranked Return Count)	`3-5 items`	Focuses on the most relevant results, reduces LLM processing load, and improves response speed.
`Max Context Length`	`3000-4000 characters`	Ensures the LLM has enough context to understand complex medical cases while avoiding exceeding limits.

Common Pitfalls

Direct chunking of the knowledge base configuration results in retrieved content that is semantically disconnected from the original document. This happens when the context dependency of medical terminology in stem cell therapy adverse reaction reports is not considered, and simple splitting destroys critical information.
A high Similarity threshold (Similarity Threshold) is set, but irrelevant quoted content is still output. This may be due to overly coarse document chunking granularity or an embedding model that does not sufficiently understand the specific semantics of the medical domain, leading to semantic drift even with high similarity.
Retrieval speed is significantly slower than expected, for example, exceeding 10 seconds. This can be caused by an excessively large knowledge base, improper indexing strategies, or insufficient hardware resources that cannot effectively handle high-concurrency retrieval requests.

Validation of Configuration

Select typical adverse event queries and compare the quoted content in the retrieval results to confirm accurate coverage and completeness of core information from the original documents.
For adverse event queries of varying severity and stem cell product types, evaluate the distribution of similarity scores to confirm high relevance of high-scoring results to the query intent.
Monitor response time under simulated concurrent query scenarios to ensure the average response time for retrieval operations meets business requirements (e.g., below 3 seconds).

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.