ARExplorer

Demo video↗️

ARExplorer -- is an agent for extracting Attitudes and Relations from documents, exposing three tools: named-entity recognition (bulk-ner), relation/attitude classification (bulk-chain), and graph set operations (union / intersection). This is the successor to ARElight, the early AREkit demo (ECIR 2024) for granular attitude/relation visualization in large documents.

Stack:

• Google ADK 2.0 — root agent, tools, session artifacts, and runtime skills (src/agent.py, src/skills/)
• bulk-ner — batch named-entity recognition (extract_named_entities)
• bulk-chain — batched LLM relation/attitude classification (classify_relations)
• Chat-driven UI — two-panel web app with d3.js graph visualization (src/static/index.html)

NER and classification backends are configured via environment variables, not baked into src/; this demo uses .recepie/arexplorer-demo/providers/ (spaCy NER + Replicate LLM adapters)

UI

Two-panel web UI — left panel chats with the agent, main panel renders the returned attitude graph with d3.js (force / radial layouts). The agent replies with a structured AgentResponse (src/schema.py); the chat shows message and the graph drives the visualization.

Usage

Either Local or Deployment.

Local

Requires Python 3.10. From the repository root:

1. Create a virtual environment and install dependencies

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Install demo provider dependencies

The NER and relation tools load adapters from .recepie/arexplorer-demo/providers/ (spaCy NER + Replicate LLM). Install their extra packages and download the spaCy model:

pip install -r .recepie/arexplorer-demo/providers/requirements.txt
python -m spacy download en_core_web_sm

3. Create .env with API keys and provider paths

Copy or create a .env file in the project root (loaded automatically by src/server.py):

cat > .env <<'EOF'
GOOGLE_API_KEY=<your-google-api-key>
REPLICATE_API_TOKEN=<your-replicate-api-token>

RELATION_MODEL=meta/meta-llama-3-70b-instruct
RELATION_PROVIDER_FILEPATH=.recepie/arexplorer-demo/providers/replicate_104.py

NER_SRC_DIR=.recepie/arexplorer-demo/providers
NER_CLASS_FILEPATH=spacy_383.py
NER_CLASS_NAME=SpacyNER
NER_MODEL=en_core_web_sm
EOF

Replace the placeholder keys before running the server.

Starting Server

uvicorn src.server:app --port 8000

Then open http://127.0.0.1:8000/.

Deployment

Create .recepie/arexplorer-demo/.env with API keys and optional auth:

cat > .recepie/arexplorer-demo/.env <<'EOF'
GOOGLE_API_KEY=<your-google-api-key>
REPLICATE_API_TOKEN=<your-replicate-api-token>
EOF

Using docker-compose:

cd .recepie/arexplorer-demo
docker compose up --build

Then open http://127.0.0.1:2000/ (Compose maps host port 2000 → container 8000).

Dependencies

• bulk-ner — batch NER over large text collections Powers extract_named_entities.
• bulk-chain — batched LLM prompting with Chain-of-Thought schemas. Powers classify_relations.