Vector Model and Indexing for Antibody-Drug Conjugate (ADC)

Data Characteristics

Antibody-Drug Conjugate (ADC) regulations and Standard Operating Procedure (SOP) data originate from internal quality management systems, R&D and production department regulations, and national drug regulatory documents. These documents are typically in PDF, Word, or internal knowledge management system formats, exhibiting high structure and standardization. Update frequency is relatively low, primarily occurring during regulatory revisions, new drug development phases, or production process optimizations. Documents contain extensive specialized terminology, chemical structure nomenclature, dosage units (e.g., mg/kg), batch information, production process steps, and quality control indicators (e.g., HPLC purity, LOD, LOQ). Document lengths range from tens to hundreds of pages, often with cross-references, forming a complex knowledge network.

Constraints Imposed by These Characteristics on Vector Models and Indexing

The highly structured and specialized terminology-intensive nature of ADC regulatory documents requires vector models to accurately capture word meaning and contextual relationships. This prevents recall bias due to ambiguity in specialized vocabulary. Cross-references between documents mean a single text chunk may not provide complete semantics. Therefore, longer context windows or multi-document association strategies are necessary. The low update frequency allows for more stable cycles for model training and index building. However, each update must ensure seamless integration of new and old knowledge. The presence of numerous numbers and units demands careful numerical representation and dimensional consistency during vectorization, especially for critical indicators like dosage and purity, to prevent information loss through generalization. Additionally, documents often include charts and chemical structural formulas, which pure text vectorization struggles to handle effectively. This may necessitate multimodal embedding or image recognition technologies.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Size)	`800–1200 characters`	ADC regulatory document paragraphs are typically long, containing complete operational steps or regulations. Longer chunk sizes help preserve contextual semantics.
`Chunk Overlap Length` (Chunk Overlap)	`100–200 characters`	Ensures sufficient contextual overlap between adjacent chunks to handle specialized terminology and process descriptions spanning multiple paragraphs.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	ADC regulatory Q&A demands high accuracy. A high threshold effectively filters out irrelevant or low-quality recall results.
`Recall count` (Recall Count)	`10–15 items`	Considering document complexity and cross-references, recalling more items helps cover potential related information and provides more comprehensive answers.
`embedding_model`	`Qwen3-Embedding-8B`	The model has strong comprehension of specialized terminology in the biomedical field, better handling ADC-related professional texts.
`maxContext`	`4096 tokens`	Ensures the model can process sufficiently long input contexts to understand complex regulatory clauses and operational procedures.

Three Common Pitfalls

Symptom: When a user asks about production regulations for a specific batch of ADC drugs, the recall results lack the critical batch management SOP. Reason: During document chunking, batch information (e.g., Batch ID) may be split across different text blocks. This leads to incomplete semantics in a single block, making precise matching difficult after vectorization.
Symptom: A Connection refused error occurs when configuring FastGPT to connect to a VLLM-deployed Qwen3-Embedding-8B model. Reason: The VLLM service is not properly started, or the listening port does not match the embedding_model connection address configured in FastGPT, preventing FastGPT from establishing a connection.
Symptom: After knowledge base construction, disk usage significantly exceeds expectations, containing many redundant files. Reason: When processing PDF and other format documents, the knowledge base may not clean up intermediate text extraction files or image caches, leading to wasted storage resources.

How to Verify Correct Configuration

Upload a typical ADC regulatory document and check the knowledge base document chunking preview. Ensure critical specialized terms, parameters (e.g., μg/mL), and operational steps are fully retained within a single text block.
Ask questions related to quality control, production processes, or compliance for specific ADC drugs. Observe the recall results to assess whether they accurately cover relevant regulatory documents. Adjust the Similarity threshold (Similarity Threshold) to observe changes in recalled items.
Check FastGPT backend logs to confirm the embedding_model connection status is normal and there are no embedding failures due to API request timeouts or parameter errors.
Regularly review the storage space usage of the vector database (e.g., Milvus). Compare actual document count and vector dimensions to assess if storage growth aligns with expectations, avoiding resource waste due to index bloat.

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.