Prove Graph Beats Tokens.

GraphRAG Inference
Hackathon

A beginner-friendly online coding challenge to prove that graphs make LLM inference faster, cheaper, and smarter. Build three pipelines side-by-side — LLM-Only, Basic RAG, GraphRAG — and let the numbers tell the story.

DatesMay 4 — May 26, 2026
Prize Pool$700 · ₹65,795
Team1 – 5 Members
WhereOnline · Global
graphrag — benchmark.pyRunning
$"Which departments had the largest revenue growth QoQ, and what product launches correlate with them?"
Pipeline 01 · LLM-Only
Just the model
Tokens12,840
Latency8.4s
Cost$0.128
Pipeline 02 · Basic RAG
Vectors + LLM
Tokens5,920
Latency3.6s
Cost$0.058
Pipeline 03 · GraphRAG
Graph + LLM
Tokens2,140
Latency1.9s
Cost$0.019
What you'll build

Three pipelines.
One question. A scoreboard that settles it.

Your mission is an interactive comparison dashboard: enter one query, run it through all three pipelines at once — LLM-Only, Basic RAG, and GraphRAG — and display side-by-side answers with tokens, latency, cost, and accuracy.

Pipeline 01

LLM-Only

Just the model

A prompt goes in, an output comes out. No retrieval — just the model reasoning from scratch. Your worst-case baseline, and the upper bound on cost.

1User promptinput
2LLM reasons over full contextexpensive
3Final answeroutput
Pipeline 02

Basic RAG

Vectors + LLM

The industry standard today. Vector embeddings find similar chunks and stuff them into the prompt. Helpful, but it can't reason across relationships between entities.

1User promptinput
2Vector search retrieves similar chunksretrieve
3Chunks stuffed into context windowstuff
4LLM answers from retrieved textanswer
Pipeline 03

GraphRAG

Graph + LLM

Built on the TigerGraph GraphRAG repo. TigerGraph handles entity extraction, relationships, and multi-hop reasoning. The LLM synthesizes from a clean, focused subgraph.

1User promptinput
2TigerGraph extracts relevant subgraphgraph
3Multi-hop reasoning over entitiesgraph
4LLM composes from structured contextcheap
5Final answeroutput
Architecture

Follow the AI Factory model.

Four separate layers — scalable, reusable, and ready for real production. A clean reference architecture for the Engineering & Storytelling axis (20% of judging).

01

Graph Layer

TigerGraph handles entity extraction, relationships, and graph queries.

02

Inference Orchestration

Decides when to use graph, when to call the LLM, and how they cooperate.

03

LLM Layer

Generates the final answer using filtered, structured context from the graph.

04

Evaluation Layer

Runs the benchmarks and populates your comparison dashboard.

Prize Pool
$700
+ TigerGraph recognition
Team Size
1 – 5
Solo allowed · cross-institutional welcome
Duration
23 Days
May 4 — May 26, 2026
Level
Beginner
No prior graph experience required
Registration closes

Every hour matters.

Registrations close May 10, 2026 · 11:59 PM IST. Don't be the team still reading docs on day one.

Who should build

Built for builders at any level.

Students

Exploring GenAI for the first time. We'll provide docs, starter code, and mentoring for the Top 10.

Developers

Curious about graph databases. If you know Python and have poked at LLMs or APIs, you're ready to build.

AI Engineers

Building production RAG. Push past vector-only retrieval — ship a graph-aware inference stack with real benchmarks.

Rewards

$700 pool. Real recognition.

All Prizes →
🏆 Winner · 1st Place
$250
Best GraphRAG Inference System
+ Certificate · TigerGraph recognition
2nd Place
$150
Runner-Up
3rd Place
$100
2nd Runner-Up
Community Award
$100
Community Leads' Choice
Content Bounty
$100
Blog · Video · Social
Timeline

From registration to the winner's mic.

Full Timeline →
Phase 01
Apr 24 — May 10
Registrations open. Join the WhatsApp group, find teammates.
Phase 02 · Round 1
May 4 — May 16
Build all three pipelines on ≥ 2M tokens. Ship the comparison dashboard. Submit on Unstop.
Phase 03 · Round 2
May 18 — May 24
Top 10 scale to 50–100M tokens with $50 Gemini credits per team + 1:1 TigerGraph mentoring.
Phase 04
May 25 — May 26
Live demos, Q&A, and winners announced.

Ship the benchmark.
Settle the question.

Show, with real numbers, exactly how much better inference gets when graphs enter the picture.