Scientific Research

Profile: scientific · Model: DEFAULT (gpt-5.6-terra) · Chunking: section · Confidence: ≥ 0.70

Overview

Parse academic papers, technical reports, and research articles to extract authors, institutions, datasets, methods, findings, and benchmark results for literature review, gap analysis, and research trend tracking.

Extracted entities

Entity type	Examples
Author	With ORCID if present
Institution	Affiliation with country
Dataset	Name, version, source URL
MLModel	Architecture, parameter count, training regime
Finding	Key result statement
Method	Experimental methodology
Metric	Benchmark result with value, dataset, and comparison

Extracted relationships

AUTHORED_BY · AFFILIATED_WITH · TRAINED_ON · EVALUATED_ON · OUTPERFORMS · REFERENCES

Section-based chunking

Academic papers have well-defined sections — abstract, introduction, methods, results, discussion, references. The scientific profile uses section-based chunking to respect these boundaries and keep related content together.

Example

curl -X POST http://localhost:8200/process \
  -H "Authorization: Bearer $API_KEY" \
  -F "file=@paper.pdf" \
  -F "profile=scientific"

{
  "entities": [
    {"type": "Author", "name": "Dr. Jane Smith", "confidence": 0.82},
    {"type": "Institution", "name": "MIT CSAIL", "confidence": 0.79},
    {"type": "MLModel", "name": "GPT-5.4", "confidence": 0.75},
    {"type": "Dataset", "name": "ImageNet-21k", "confidence": 0.77},
    {"type": "Metric", "name": "Top-1 accuracy: 92.3%", "confidence": 0.73},
    {"type": "Finding", "name": "Scaling to 500B parameters yields diminishing returns on reasoning tasks", "confidence": 0.70}
  ],
  "relationships": [
    {"source": "Dr. Jane Smith", "target": "MIT CSAIL", "type": "AFFILIATED_WITH"},
    {"source": "GPT-5.4", "target": "ImageNet-21k", "type": "EVALUATED_ON"}
  ]
}

Typical workflow

arXiv papers / internal reports (PDF)
    ↓
DocIntel (scientific profile) → authors, methods, findings, metrics
    ↓
Purple8 Graph → research knowledge graph
    ↓
Cortex → literature review, gap analysis, trend queries

Multi-language support

Scientific papers in non-English languages benefit from multi-language extraction. Entity names are extracted in native script with English JSON keys. 75+ languages supported.