Data Characteristics
Stem cell therapy data originates from clinical trial reports, drug submission documents, academic research papers, patent literature, and regulatory guidelines. These documents update infrequently, typically aligning with clinical trial cycles, approval processes, or research advancements. Annual or quarterly updates are common. Documents are often long, unstructured, or semi-structured texts. They contain specialized terminology, experimental data, charts, references, and complex biological pathway descriptions. Fields include cell types (e.g., mesenchymal stem cells, hematopoietic stem cells), disease indications, administration routes, dosages, efficacy metrics (e.g., CR, PR), safety data (e.g., SAE incidence), mechanisms of action, and manufacturing process details. Units vary, including cells, mL, cells/mL, days, weeks, months, and %.
Constraints on Knowledge Base Retrieval and Recall
The specialized and complex nature of stem cell therapy data presents several challenges for knowledge base retrieval and recall. First, long documents and dense technical details require fine-grained text segmentation to avoid context loss or information fragmentation. Second, the high density of specialized terms and abbreviations demands that tokenizers and embedding models have strong domain understanding to accurately capture semantics. Third, data updates are infrequent, but each update may contain critical clinical results or approval progress. The retrieval system must index the latest versions promptly and differentiate between versions. Fourth, diverse units and numerical data mean simple keyword matching is insufficient. Numerical range search or unit conversion capabilities may be necessary. Finally, for critical safety indicators like SAE incidence, users often focus on specific values. Recall results must pinpoint relevant paragraphs and extract the values themselves.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 800–1200 characters (characters) | Balances context completeness and retrieval granularity for long professional texts. |
Chunk Overlap Length (Segment Overlap Length) | 100 characters (characters) | Ensures contextual continuity between segments and reduces semantic fragmentation from splitting. |
Recall count (Number of Retrieved Items) | Top 10 entries (top 10) | Provides more potentially relevant information for subsequent re-ranking or LLM filtering, given the complexity of stem cell therapy. |
Similarity threshold (Similarity Threshold) | Calibrate by actual measurement (calibrate based on actual measurements) | Requires adjustment based on the specific embedding model and dataset to balance recall and precision. |
Rerank result count (Number of Re-ranked Items) | Top 5 entries (top 5) | Provides the most relevant and diverse results after optimization by the re-ranking model. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Allows sufficient parsing time for large PDF or DOCX documents. |
Common Pitfalls
- Uploading
xlsxclinical data tables to the knowledge base results in the model being unable to read their content. The default text extractor may not effectively parse complex table structures into retrievable plain text, leading to unindexed table data. - Retrieving information on specific stem cell product dosages and administration routes returns general results, failing to pinpoint precise values or protocols. The segmentation strategy may be too coarse, mixing critical numerical information with irrelevant content and diluting key information in the embedding vectors.
- A user query about a newly approved stem cell therapy returns outdated clinical trial data. The knowledge base's document version management may be inadequate, failing to prioritize new versions or replace old ones, leading to retrieval of obsolete information.
Validation Steps
- Upload a
PDFdocument containing stem cell product clinical trial results. Search using keywords like "SAEIncidence" (SAE incidence) or "dosage5x10^6 cells/kg" (dosage 5x10^6 cells/kg). Verify that the results accurately hit paragraphs containing these specific values and units. - Select several representative stem cell research papers. Upload them to the knowledge base. Ask questions about core biological pathways or mechanisms of action. Observe if the model recalls paragraphs containing these specialized terms and evaluate the semantic relevance of the returned paragraphs.
- Simulate a user query for the latest approval status or updates of a stem cell therapy. Check if the recall results include the most recently published documents or document fragments with the latest update dates. If multiple versions exist, ensure the latest version is prioritized.
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.