Agentic GraphRAG Hackathon

TigerGraph Agentic GraphRAG HackathonSome questions can be answered with a single retrieval step. But the hard ones — the ones that actually matter — need a system that can break the question apart, figure out what to retrieve, evaluate what comes back, spot gaps, and keep digging until it has a real answer.This hackathon is about building that. You'll create an Agentic GraphRAG system on TigerGraph that autonomously investigates complex questions using graph, vector, and document evidence. The goal isn't to prove agentic is always better; it's to figure out exactly where it matters and where simpler approaches are enough.Open to college students and working professionals. Solo participants or teams of up to 5 are allowed. Free to register.Click here to join the WhatsApp group: Click HereWhy This Is HardA basic RAG pipeline retrieves some text and generates an answer. That works for simple questions. GraphRAG adds structure — entities, relationships, community context — and that works for harder ones.But some questions need the system to actually think. Break the problem down. Decide what to look up. Evaluate what came back. Realize something's missing. Go back and retrieve more. Change strategy mid-investigation. Decide when it has enough to answer. That's the agentic part, and that's where this challenge lives.The ChallengeCore (everyone builds this): An Agentic GraphRAG system that can plan and execute a multi-step investigation. Your system should include:An agent harness to manage state, tools, context, evidence, and stopping criteriaAn orchestrator agent that determines what needs investigating and picks the next action — not a fixed retrieval sequenceSpecialized agents for retrieval and reasoning tasks: entity linking, graph traversal, similarity search, document retrieval, aggregation, multi-hop reasoning, and evidence evaluationThe orchestrator's next move should depend on the original question, the graph, what evidence it already has, and what's still missing.The benchmark: Every question gets answered three ways — RAG, GraphRAG, and Agentic GraphRAG — so we can see exactly where each approach succeeds or fails.Stretch (for teams who want to go further): Reasoning over time. Real-world facts change — launch dates move, plans get revised, and sources conflict. Extend your agent to detect conflicting versions of a fact, figure out what supersedes what, determine which sources are more authoritative, and handle uncertainty. This is where strong teams pull ahead.Two ground rules:Build the benchmark. Your solution must compare all three approaches (RAG, GraphRAG, Agentic GraphRAG) using accuracy, completeness, and token efficiency metrics. That comparison is the point.Add-ons are welcome. Beyond the core and stretch, build whatever makes your system smarter or more useful. Creative add-ons count in your favour under engineering and innovation.What You Build WithSavanna: Host the graph and vector DB (free credits provided)GSQL + graph algorithms: Traversal, pattern matching, community detection, and investigation logicTigerGraph Vector DB: Semantic search across unstructured documentsTigerGraph MCP: Connect agents and development assistants to TigerGraphDataset: We provide a dataset with questions of increasing complexity. You're free to bring your own too. The provided dataset is the common benchmark everyone is scored on, so use your own as a bonus to show your system can handle real or messy data.How the Two Rounds WorkRound 1 (open to everyone): Submit a working Agentic GraphRAG system covering the core challenge.Round 2 (top 15 teams advance): Extend the agent to reason over evolving, conflicting, and uncertain facts on a harder dataset. Finalists also submit a demo video and write-up. Top teams present live to the judging panel.TimelineSeptember 1 (Tue): Registration opens; guidebook and dataset go liveSeptember 12 (Sat): Registration closesSeptember 24 (Thu): Round 1 submission deadlineOctober 1 (Thu): Round 2 / final submission deadlineOctober 2 to 5 (Fri to Mon): JudgingOctober 7 (Wed): Results announcedWhat to SubmitRound 1: Working Agentic GraphRAG system + GitHub repository + architecture diagram + demo video + metrics dashboard (tokens, accuracy, and completeness across all three pipelines). Optional: A social media post about what you're building (counts in your favour).Round 2 (finalists): Refined system + updated repository + architecture diagram + 3–5 minute demo video + metrics dashboard + short write-up (what you built, how it works, key results, limitations, and what you'd do with more time).How It's JudgedAccuracy and completeness are scored against a held-out ground-truth dataset. Quality is assessed via automated evaluation, LLM-as-judge, human review, and repository review.Investigation accuracy, 30%: Answers complex questions correctly using the right evidence across retrieval and reasoning stepsEvidence quality and explainability, 15%: Grounded answers with clear citations and a clear investigation pathAgentic effectiveness and efficiency, 15%: Picks the right retrieval methods, uses agentic steps where they add value, and balances accuracy with token costAgentic design, engineering, and code quality, 15%: Architecture, tool use, reliability, reproducibility, and repository qualityInnovation, 15%: Novel investigation methods, graph reasoning, or user experienceFinal presentation and Q&A, 10%: Demo quality, technical clarity, and responses to judgesPrizesTotal prize pool: ₹70,0001st place: ₹35,0002nd place: ₹20,0003rd place: ₹15,000Certificates for all valid submissions. Top teams featured across TigerGraph's community channels.