Data Characteristics
Antibody-Drug Conjugate (ADC) R&D documents originate from various sources. These include early target discovery reports, preclinical study data (in vitro and in vivo efficacy, toxicology reports), clinical trial protocols and results, manufacturing process records, and quality control standards. Document update frequency is high, especially during clinical trials, where data generation is continuous. Document structures range from highly structured experimental data tables to semi-structured research reports (containing figures, tables, and text descriptions), and unstructured meeting minutes or expert review opinions. Specific fields and units for ADC drugs include molecular structure information for antibodies, linkers, and payloads, as well as parameters like drug-antibody ratio (DAR value), pharmacokinetic/pharmacodynamic (PK/PD) parameters, toxicity indicators, and biomarkers. These involve various specialized units such as molar concentration, dosage (mg/kg), time (h), and biological activity units (nM, μg/mL).
Constraints on Knowledge Base Retrieval and Recall
The data characteristics of ADC R&D documents impose specific requirements on knowledge base retrieval and recall. Diverse data sources and high update frequency necessitate support for rapid import of multiple file formats, incremental updates, and effective version management. The coexistence of structured and unstructured data requires a chunking strategy that differentiates text paragraphs, tabular data, and figure captions. This prevents critical information from being fragmented or confused. ADC-specific biomolecular structures, DAR values, and PK/PD parameters require vector models with deep understanding of biomedical terminology. This ensures accurate semantic matching for similar concepts expressed differently. Diverse unit systems demand identification and normalization during information extraction and comparison, preventing recall failures due to unit discrepancies. For example, retrieving DAR values may involve multiple expressions, requiring the model to recognize and associate them.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 500 MB | ADC R&D documents often contain large experimental datasets and high-resolution images, leading to larger file sizes. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Complex PDF or Word documents require sufficient parsing time. |
Chunk size (Chunk Length) | 800–1200 characters | Balances contextual coherence for long reports with the need for fine-grained retrieval, avoiding excessive fragmentation. |
Recall count (Recall Count) | 10 entries | Ensures coverage of multiple relevant chunks, especially for complex queries. |
Similarity threshold (Similarity Threshold) | 0.75 | ADC specialized terminology has high semantic distinctiveness; increasing the threshold reduces irrelevant recalls. |
Rerank result count (Reranked Return Count) | Top 5 entries | Further enhances core relevance using a reranking model based on initial recall. |
Common Pitfalls
- Knowledge base file upload error
fail to create post presigned url: This typically results from misconfigurations in object storage services like S3 or MinIO, such as insufficient permissions or expired access credentials, preventing FastGPT from obtaining a presigned URL for file uploads. - Missing or incomplete critical experimental data in search results: This indicates an improper chunking strategy. For example, fragmenting paragraphs containing tabular data too finely can prevent individual chunks from providing complete context.
- Inaccurate retrieval results for "Conjugation Ratio" (conjugation ratio) or "DAR值" (DAR value): The vector model may lack sufficient understanding of specific biomedical terminology. It fails to associate synonymous but differently expressed professional terms, affecting semantic matching.
Verification Steps
- Upload various types of ADC R&D documents (PDF, Word, Excel). Confirm successful parsing and ingestion for all files, then check their document status.
- Perform retrieval operations using specific ADC molecule names, targets, and DAR value ranges as professional terms. Check if the recalled results include critical information segments.
- Select a toxicology report containing complex tables. Retrieve a specific data point (e.g., body weight change for a particular dose group). Confirm that the recalled chunk completely presents the relevant table row or column.
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.