title	emoji	colorFrom	colorTo	sdk	app_port	pinned
Uncharted Waters	🌊	gray	yellow	docker	7860	false

Uncharted Waters

Automated research landscape analysis against the DTIC Dimensions database. Describe a research area and get back a structured assessment of what already exists, what's missing, and where the opportunities are.

Try it live on HuggingFace Spaces

How It Works

The pipeline runs four stages:

1. Search — Multiple query strategies scan the DTIC Dimensions publication database, deduplicating across result sets
2. Embed — Publications and the research topic are encoded with nomic-embed-text-v1.5 using asymmetric retrieval prefixes (search_query: / search_document:)
3. Score — Similarity is computed as the geometric mean of holistic embedding similarity and IDF-weighted per-keyword concept scores. Keywords that appear in many results (generic terms) are down-weighted; rare, specific keywords carry more signal. This prevents a general survey paper from inflating overlap when the research topic is a specific multi-concept intersection
4. Assess — Claude analyzes the scored results and generates a landscape report with comparisons, gaps, and recommendations. The verdict and confidence are computed deterministically from scores and branch data — the LLM provides narrative, not metrics

Verdicts

Verdict	Meaning
Open Landscape	No substantially similar work found
Branch Opportunity	Similar work exists but funded by other branches
Well Covered	Very similar existing work found in the same branch
Mixed Coverage	Partial overlap — requires expert judgment

Development

# Clone and install (CPU-only PyTorch saves ~1.5GB)
git clone https://github.com/that-github-user/uncharted-waters.git
cd uncharted-waters
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements-dev.txt

# Configure
cp .env.example .env
# Edit .env — set ANTHROPIC_API_KEY

# Run locally
uvicorn src.api:app --reload
# → http://localhost:8000

# Run tests (all mock HTTP, no API key needed)
pytest tests/ -v

Cost

Each analysis run uses approximately $0.02–0.05 in Anthropic API usage (Claude Sonnet).