Domain-Aware Extraction
Extract chemical structures, entities, and relationships from unstructured documents.
/ The problem
Chemical structures in research documents are images — invisible to traditional text search. Compound identifiers like SACC-3060 or BMS-986158 are opaque to general-purpose embeddings. Named entities (genes, targets, assays) and their relationships are buried in prose, lost to any system that only indexes raw text.
/ How DocuStore solves it
DocuStore combines Optical Chemical Structure Recognition (OCSR) with hybrid Named Entity Recognition to extract domain structure from every page. The OCSR pipeline uses YOLO-based structure detection, neural template matching, and DECIMER to convert molecular diagrams into canonical SMILES strings. The NER system runs in dual mode: dictionary-based for exact matches (compounds, genes, targets) and LLM-powered for contextual extraction (diseases, measurements, compound-target-assay relationships).
/ Pipeline
PDF Page Image
↓ YOLO structure detection
Molecular Region Crops
↓ Neural template matching
Matched / Unmatched Structures
↓ DECIMER SMILES extraction
Canonical SMILES + Confidence Score
↓ Parallel: Dictionary NER + LLM NER
Entities: compounds, genes, targets, assays, measurements
↓ Relationship extraction
Compound-Target-Assay-Value records/ Key capabilities
- OCSR via structflo-cser: YOLO detector + DECIMER SMILES extraction
- Hybrid NER: dictionary-based exact matching + LLM contextual extraction
- Multi-level summarization: page-level (text+image hybrid) and document-level
- Relationship extraction: compound-target-assay triples with quantitative values
- Automatic workflow orchestration via Temporal with retry and error handling
- Configurable LLM providers: Ollama (local), OpenAI, or Gemini