TigerGraph Agentic GraphRAG Hackathon Some 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. Join the community for questions, debugging help, and office hours during the build window. All important updates - guidebook, dataset releases, deadline reminders, and announcements - will be shared here first:
Discord: https://discord.gg/eKWm3mbkw2
WhatsApp Group: https://chat.whatsapp.com/GN0x6yajuroDJTzLMEZvBR
Guidebook: https://alluring-beryllium-491.notion.site/Agentic-GraphRAG-Hackathon-Guidebook-34fc2cb129c08146998af3568d7d2594
Dataset: https://drive.google.com/drive/folders/10C0hzRaHlm00VYPFbjapKtWj0EPmLvQ9?usp=drive_link Why This Is Hard A 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 Challenge
Core (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 criteria An orchestrator agent that determines what needs investigating and picks the next action — not a fixed retrieval sequence Specialized agents for retrieval and reasoning tasks: entity linking, graph traversal, similarity search, document retrieval, aggregation, multi-hop reasoning, and evidence evaluation The 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
With Savanna: Host the graph and vector DB (free credits provided) GSQL + graph algorithms: Traversal, pattern matching, community detection, and investigation logic TigerGraph
Vector DB: Semantic search across unstructured documents TigerGraph
MCP: Connect agents and development assistants to
TigerGraph Dataset: 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 Work Round 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.
Timeline September 1 (Tue): Registration opens; guidebook and dataset go live
September 12 (Sat): Registration closes
September 24 (Thu): Round 1 submission deadline
October 1 (Thu): Round 2 / final submission deadline October 2 to 5 (Fri to
Judging October 7 (Wed): Results announced What to Submit
Round 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 + m