GraphRAG Developer Challenge – Legal Document Processing (Prototype)
Role Details

Role: Senior RAG Systems Developer (Contract / Freelance)
Compensation: $600 — paid only if you pass (95% benchmark)
Timeline: 3–5 days from materials receipt to live demo
Purpose: Technical evaluation for potential long-term hire
Frontend / UI: None (backend prototype only)

Objective
We’re seeking an expert in graph-based retrieval (GraphRAG) to build a high-accuracy prototype for legal document reasoning. This is a paid technical test that may lead to a long-term position. The goal is a true GraphRAG system featuring explicit knowledge-graph construction and traversal, multi-hop reasoning, agentic orchestration, and strong focus on retrieval accuracy and explainability.
Materials Provided

/docs/ → Pre-processed Markdown legal documents with metadata
/sample_questions.json → Example question format
/sample_answers_rag.json → Example answer format

Download materials:
(Benchmark uses unseen questions.)
Deliverables
Implement two functions in Python 3.12 (Poetry project):

def ingest(document_paths: List[str]) -> None: 'Ingest Markdown docs and build the knowledge graph.'
def query(questions: List[str]) -> List[str]: 'Return answers with Vancouver-style citations grounded in retrieved sources.'

Requirements: No UI, no API keys provided. Any stack may be used. query(...) must support parallel execution (~400 questions in ≤60 min) and show a progress indicator. Test thoroughly for correctness and performance before the demo.
Live Demo
In a 60-minute live session you will:

Receive ~400 unseen questions.
Run query(...) to produce answers.json.
Explain your architecture: how the graph is built, traversed, and used to generate grounded answers.
Only the developer(s) who wrote the code may present.

Evaluation
Passing requires an overall score above 95%, measured by (LLM as a judge): Faithfulness (grounded, no hallucinations), Relevance (retrieval matches intent), Completeness (covers key legal points), and Clarity (structured, legally coherent writing).
#J-18808-Ljbffr