CAR-T Cell Therapy R&D Document Structuring: Citation and

Data Characteristics

CAR-T cell therapy R&D documents come from various sources. These include clinical trial protocols, investigator brochures, patient informed consent forms, pharmaceutical/non-clinical study reports (e.g., pharmacokinetic, pharmacodynamic, toxicology reports), and regulatory submission documents. Document updates are closely tied to the R&D phase; early research sees frequent updates, while late-stage clinical development is relatively stable. Document structures are complex, often containing extensive unstructured text, tables, and graphs. Fields and units are highly specialized. For example, dose units often involve cells/kg or ug/mL. Time units are precise, such as days, weeks, or months. Many specialized biological terms and acronyms are present, like DLT (dose-limiting toxicity) and CR (complete remission).

Constraints on Citation and Traceability

The complexity of CAR-T cell therapy R&D documents imposes specific requirements on citation and traceability. First, dense specialized terminology and acronyms can lead to ambiguity during model recall, affecting traceability accuracy. This requires more refined text segmentation strategies. Second, critical data in tables and graphs are often core citation points. Traditional text-block-based recall methods struggle to locate these directly, necessitating enhanced parsing capabilities for table content or graph descriptions. Third, varying document update frequencies across R&D phases require the knowledge base to manage versions effectively, ensuring citations refer to the latest or specified versions. Finally, strict regulatory requirements mean any cited content must be traceable with high precision, pointing to specific page numbers or sections in the original document to meet audit and compliance needs.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters`	Balances contextual completeness of long paragraphs in CAR-T documents with retrieval efficiency, avoiding excessive truncation of key information.
`Chunk Overlap Length` (Segment Overlap Length)	`100–150 characters`	Ensures semantic coherence between adjacent segments, especially at transitions involving specialized terms and data descriptions.
`Recall count` (Recall Count)	`8–12 entries`	Covers potentially dispersed key information points in CAR-T R&D documents while maintaining recall relevance.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Balances precise matching of specialized terms with semantic similarity, reducing interference from irrelevant results.
`Rerank result count` (Reranked Return Count)	`3–5 entries`	Focuses on core evidence most relevant to specific CAR-T questions, reducing redundant information.
`ENABLE_RAG_REFERENCE_AUTO_EXTRACT`	`true`	Ensures automatic extraction and display of citation sources from model answers, meeting traceability requirements.

Common Pitfalls

Cited knowledge base entries in model answers are empty or do not match the actual answer content. This happens when the knowledge base segmentation strategy is inadequate, failing to capture critical information in CAR-T R&D documents. Alternatively, the similarity threshold for recall might be too high, filtering out relevant segments.
After a workflow tool call, the answer still displays input and response citation information from the knowledge base search. This occurs when the Output Settings in the Workflow Configuration are not correctly configured. Citation display rules might not be disabled or adjusted as needed, causing default output formats to persist.
Critical data in specific tables or graphs cannot be cited or traced. This happens when the knowledge base's document parser fails to effectively identify and extract non-textual data structures, leading to their omission during vectorization and retrieval.

Verification Steps

Ask typical CAR-T R&D questions. Check if the citation sources in the answer accurately point to specific sections or paragraphs in the original document.
Verify that the cited content in the answer is identical to the phrasing in the original document, especially for critical data like dosages, time points, and clinical results.
Simulate document updates. Test if the knowledge base's answers correctly cite the latest version of the file and ensure old version citations no longer appear or are clearly identified.
Check that input or response citation information, which should not be displayed after a workflow tool call, does not appear in the answer. Confirm Output Settings are effective.

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.