CogniDoc

CogniDoc is an open-source workspace for building agentic document intelligence systems. Its goal is to turn complex, unstructured documents into traceable, structured data by combining document curation, visual parsing, schema-aware extraction, and human inspection.

The repository currently includes a runnable Data Studio for managing machine-learning datasets, plus research and early components for OCR, table recognition, and document extraction.

CogniDoc is under active development. The Data Studio is the most complete application today; the curator, parser, and extractor are evolving research components.

Why CogniDoc?

Real-world documents are more than streams of text. Tables, reading order, figures, page boundaries, and visual position all carry meaning. CogniDoc is designed around a pipeline that preserves this structure:

flowchart LR
    A[Documents] --> B[Curator]
    B --> C[Parser]
    C --> D[Extractor]
    D --> E[Structured JSON]

    C --> F[(Session data)]
    D --> F
    F --> G[Review and evaluation]

• Curator prepares documents by splitting, cleaning, classifying, and grouping pages.
• Parser reconstructs text, tables, figures, layout, and reading order while retaining visual grounding.
• Extractor maps parsed evidence to a target schema, validates candidate values, and produces structured output.
• Data Studio stores, versions, previews, and shares the datasets used to develop and evaluate the pipeline.

What is available

Data Studio highlights

The included TonAI Data Studio accepts common Hugging Face Dataset Hub repository layouts without reorganizing their files. It provides:

• folder uploads with safe relative-path preservation;
• Dataset Card rendering from README.md;
• Parquet, CSV, TSV, JSON, JSONL, TXT, and ImageFolder discovery;
• dataset configs, splits, schemas, statistics, and bounded row previews;
• immutable, content-addressed revisions backed by Git and DVC;
• local or S3-compatible object storage;
• public, internal, and private dataset visibility;
• a FastAPI backend and React web interface; and
• a Docker Compose stack with PostgreSQL, Redis, Celery, and RustFS.

Quick start

The quickest way to run the current application is with Docker Compose.

Requirements

• Docker with the Compose plugin
• Git

Start the stack

cd src/data_studio
cp .env.example .env

Replace the placeholder secrets in .env, then run:

docker compose --env-file .env \
  -f infrastructure/docker-compose.yml \
  up -d --build

Once the services are healthy, open:

• Web interface: http://localhost:3000
• API documentation: http://localhost:8001/docs
• RustFS console: http://localhost:9001/rustfs/console/

The first account registered in the Studio becomes the workspace administrator.

Stop the services without deleting their persistent volumes:

docker compose --env-file .env \
  -f infrastructure/docker-compose.yml \
  down

For direct Python and Node.js development, configuration options, API examples, and the complete feature list, see the Data Studio documentation.

Repository structure

cognidoc/
├── src/
│   ├── curator/               # Document preparation (planned)
│   ├── parser/                # Experimental OCR tool adapters
│   ├── extractor/             # Structured extraction (planned)
│   └── data_studio/           # Runnable dataset management application
├── research/
│   ├── chandra2-parser/       # Chandra OCR experiments
│   ├── table_cell_detection/  # Table and cell detection experiments
│   ├── table_html_dataset/    # Table-to-HTML dataset tooling
│   └── table_recognition_vlm/ # SFT and RLVR training workflows
├── docs/                      # Architecture and design notes
└── asssets/                   # Samples and research assets

Development

Data Studio uses Python 3.11+, FastAPI, SQLAlchemy, React, TypeScript, and Vite. Direct development requires Python 3.11+, Node.js 22+, and uv. From src/data_studio, install its development dependencies:

uv pip install -e '.[dev]'
cd apps/web
npm ci
cd ../..

Run the backend checks:

ruff check apps/api tests migrations
ruff format --check apps/api tests migrations
mypy apps/api/data_studio_api
pytest

Run the frontend checks:

cd apps/web
npm run lint
npm test
npm run build

The research directories have separate dependencies and instructions. Start with the table-recognition SFT guide or the SFT + RLVR guide.

Documentation

• Project vision and document-processing architecture
• Data Studio guide
• Data Studio API examples
• Session database design
• English product documentation
• Vietnamese product documentation

License

CogniDoc is available under the MIT License.