Data Characteristics for This Category
CAR-T cell therapy registration dossiers primarily include preclinical study data, clinical trial data, manufacturing process information, and quality control standards. Data sources are diverse, encompassing in-vitro experiment reports, animal experiment reports, Phase I/II/III clinical trial reports, and manufacturing batch records. Data update frequency is high during clinical trial phases, especially for safety and efficacy data, which may update quarterly or semi-annually. Document structures are complex, often containing PDF reports, Excel statistical data, and Word protocols and summaries. Fields and units are highly specialized, such as "cell expansion fold" (X-fold), "CD3+ cell percentage" (%), "viral vector copy number" (VCG/cell), and "adverse event grade" (CTCAE v5.0 grade). Data from different sources may also have inconsistent units or naming conventions.
Constraints Imposed by These Characteristics on Tool Calling and Plugins
The diversity of CAR-T cell therapy data sources and complex document structures require tool calling and plugins to have robust file parsing and heterogeneous data integration capabilities. For example, extracting key safety data from PDF reports or reading specific batch quality control indicators from Excel spreadsheets requires precise text recognition and structured extraction. High-frequency data updates, especially for clinical trial data, demand that tools support data synchronization and incremental processing to avoid reprocessing large existing datasets. Specialized fields and units, along with potential naming differences, necessitate strict validation and standardized conversion during parameter passing and result parsing. Failure to do so can lead to data misalignment or calculation errors. Furthermore, due to the rigorous nature of registration dossiers, tools must provide clear error messages and traceability mechanisms when calls fail.
Configuration Settings
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
maxContext | 4000 | To handle complex reports and multi-source data integration, maintaining context coherence. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | To process large PDF reports or complex Excel statistical tables, preventing parsing timeouts. |
Similarity threshold | 0.75 | To ensure retrieved regulatory provisions or previous submission cases are highly relevant to CAR-T data. |
Rerank result count | Top 5 entries | To select the most relevant items from a large number of search results, improving subsequent processing efficiency. |
mcp_tool_retries | 3 | To handle external MCP tool call failures due to network fluctuations or temporary service unavailability. |
enable_streaming_output | no | Responses for registration dossiers typically require completeness and accuracy. Disabling streaming output ensures the full presentation of final results. |
Three Common Pitfalls
- Tool calls return empty or incomplete data because the PDF parser fails to correctly identify table structures or key fields in the report.
- API call text output speed is too slow because the backend service takes a long time to process large amounts of structured data extraction and complex logical judgments, without optimized asynchronous processing.
- Tool call records auto-reply, but many MCP tools cannot connect to the call termination node. This occurs because the MCP tool's output format does not match expectations or lacks necessary interactive protocol support.
How to Verify Correct Configuration
- Use test cases to verify that tools accurately parse different formats of CAR-T submission documents and successfully extract key fields. Check the accuracy of extracted values and unit consistency.
- Execute a simulated submission process. Observe the response time and data processing speed of the entire toolchain. Ensure API output speed meets requirements and compare it against baseline times.
- Check tool call logs to confirm that all MCP tool call statuses are successful and that the call termination node correctly receives and processes tool return results.
The values provided are common starting points and should be measured 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.