Vector Model and Indexing for CAR-T Cell Therapy Regulations

Data Characteristics

CAR-T cell therapy regulatory and SOP documents typically originate from pharmaceutical regulatory bodies, internal clinical trial protocols, and hospital quality management system files. These documents exist as PDFs, Word files, or internal knowledge base pages. Their content is structured, containing specialized terminology, operating procedures, quality control standards, and adverse reaction handling processes. Regulatory documents update infrequently, usually annually or with policy changes. Internal clinical SOPs might revise quarterly or semi-annually based on technological advancements or practical experience. Documents often include precise numbers and units, such as dosage (10^6 cells/kg), timeframes (28 days follow-up), and temperature requirements (-150℃ storage).

Constraints on Vector Models and Indexing

The specialized and rigorous nature of CAR-T cell therapy regulatory documents requires vector models to accurately understand and differentiate synonyms, near-synonyms, and context-specific meanings. This avoids recall errors due to semantic drift. The precise numbers, units, and operating procedures in documents challenge chunking strategies. Key information must remain intact, and paragraph semantic integrity must be maintained. Varying update frequencies require indexing mechanisms to support incremental updates, efficiently incorporating revisions without full rebuilds. Furthermore, regulatory compliance mandates high recall and accuracy. This means similarity thresholds and reranking strategies need careful configuration to ensure answers align with regulations. For multilingual or specific terminology, a pre-trained model's domain adaptability is also an important consideration.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`chunk_size`	`500–800 characters`	Ensures semantic integrity of individual operating steps or regulatory clauses, preventing truncation of key information.
`overlap_size`	`100–150 characters`	Retains contextual connections, helping the vector model understand cross-paragraph references.
`embedding_model`	`qwen3-embedding-8b` or `m3e`	Selects models that perform well on Chinese corpora for Chinese biomedical terminology.
`similarity_threshold`	`0.75–0.85`	Improves answer relevance and accuracy while maintaining recall, reducing interference from irrelevant information.
`recall_count`	`8–12 items`	Given the complexity of CAR-T regulations, increases recall to cover more potentially relevant passages.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Accommodates parsing time for large regulatory files or multi-page SOP documents, preventing timeout interruptions.

Common Pitfalls

Files remain in an "indexing" state after upload and do not complete. This often happens because the selected embedding_model is unsupported or fails to load, causing the vectorization process to not start or terminate abnormally.
After a user query, the returned answer lacks critical numerical or unit information, or contains incorrect operating steps. This often results from chunk_size being set too large or too small, causing key information to be split or mixed with irrelevant content, affecting model understanding.
For specific terminology or abbreviations, the model cannot accurately recall relevant document snippets. This may indicate insufficient domain adaptability of the embedding_model, with limited understanding of biomedical professional vocabulary, requiring consideration of fine-tuning or model replacement.

Validation Steps

Upload representative CAR-T regulatory documents. Observe if the file status changes to "completed" within a reasonable time. Check the logs for any abnormal errors.
Ask questions about specific operating steps, dosage requirements, or adverse reaction handling processes within the document. Verify if the returned answers are accurate, complete, and include all key numbers and units.
Use uncommon professional terms or abbreviations from the document for retrieval. Evaluate if the recall results include relevant document snippets. Observe changes in recall effectiveness by adjusting the similarity_threshold.
After updating documents, perform incremental indexing. Confirm that new content is quickly incorporated and participates in retrieval without affecting the availability of older content.

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.