KankerWijzer

KankerWijzer is a provenance-first cancer information assistant built on trusted IKNL-aligned sources. It combines structured registry/statistics APIs, trusted content retrieval, safety guardrails, and a FastAPI backend so questions can be answered with source-aware evidence instead of generic LLM guesses.

Warning

Current status: pre-production open-source prototype. This repository is tested and developer-ready, but it is not yet suitable for unsupervised clinical use or medical decision-making.

What Is Included

• FastAPI backend for retrieval, answering, feedback collection, and Lastmeter resources
• Trusted-source adapters for kanker.nl, nkr-cijfers, kankeratlas, richtlijnendatabase, IKNL reports, and scientific publications
• Structured-first routing for statistics and regional questions before vector search
• Guardrails for emergency language, distress routing, source whitelisting, and evidence abstention
• Local evaluation prompts and regression tests for unsafe prompt handling
• Docker Compose stack for Postgres, Qdrant, and Neo4j

Repository Layout

.
├── backend/                  FastAPI app, retrieval layer, tests, evaluation scripts
├── frontend/                 Static UI prototype
├── artifacts/                Evaluation and demo artifacts
├── problem-statement/        Copied hackathon brief and source context
├── docker-compose.yml        Local infra stack
├── CHANGELOG.md              Release notes
├── CONTRIBUTING.md           Contributor workflow
├── SECURITY.md               Security and safety reporting process
└── CODE_OF_CONDUCT.md        Community expectations

Quickstart

1. Start the backend

cd backend
uv sync --extra dev
uv run uvicorn app.main:app --reload

2. Start local infrastructure

docker compose up -d postgres qdrant neo4j

3. Run the test suite

cd backend
python3 -m pytest -q

Safety Model

• Approved-source whitelist only
• Refusal on diagnosis-seeking and treatment-decision prompts
• Red-flag routing for crisis and emergency language
• Evidence-threshold abstention when retrieval confidence is too low
• Distress detection with Lastmeter resource routing
• Provenance attached to structured and retrieved evidence

Production Readiness

Improved in this repo

• Startup now degrades gracefully when optional infrastructure dependencies such as Qdrant or Neo4j are absent
• CI-ready import and API smoke coverage has been added
• Atlas routing now uses query-specific cancer group, sex, and postcode granularity instead of a hardcoded sample response
• Open-source project hygiene now includes changelog, contributing guide, code of conduct, security policy, templates, and CI workflow

Still needed before calling this production-grade

• Clinical validation with domain experts and documented acceptance criteria
• Authentication, authorization, audit logging, and rate limiting
• Production observability: metrics, tracing, alerting, SLOs, and incident response
• Secret management and hardened deployment configuration
• Broader query-driven guideline coverage and more extensive end-to-end evaluation on real user journeys

Development Workflow

• See CONTRIBUTING.md for setup and pull request expectations.
• See CHANGELOG.md for tracked changes.
• See SECURITY.md for responsible disclosure and medical-safety escalation guidance.

Demo and Supporting Material

• PRESENTATION.md
• PLAN.md
• backend/README.md

License

This project inherits the hackathon repository's licensing context unless and until the maintainers publish a separate license for the standalone repository.