![[DSC Europe 23] Djordje Grozdic - Transforming Business Process Automation wi...](https://cdn.slidesharecdn.com/ss_thumbnails/djordjegrozdic-transformingbusinessprocessautomationwithretrieval-augmentedgenerationandllms-231129100159-8fe5bf2b-thumbnail.jpg?width=560&fit=bounds)
Ad
More Related Content
Similar to Scaling GraphRAG: Efficient Knowledge Retrieval for Enterprise AI (20)
Recently uploaded (20)
![Minitab 22 Full Crack Plus Product Key Free Download [Latest] 2025](https://cdn.slidesharecdn.com/ss_thumbnails/avtpastoralvisits20252026-250407114051-0c8f0fbf-250407123507-bf8f5b04-250426175635-00bc6197-thumbnail.jpg?width=560&fit=bounds)
Ad
Scaling GraphRAG: Efficient Knowledge Retrieval for Enterprise AI
- 1. Scaling GraphRAG: Efficient Knowledge Retrieval for Enterprise AI DevOps X AWS Startup Circuit | April, 2025 Guy Korland, PhD. CEO & Co-Founder FalkorDB Accurate, Coherent, & Traceable Retrieval for LLMs
- 3. “Parent” split “Parent” split P1 P2 Child Child Children Child Child Children Child Child Children Child Child Children Child Child Children Child Child Children Small chunks stored in Vectorestore Large chunks stored as key-value pairs in ‘InMemorystore’ VectorDB <> LLMs Naive, lacks domain awareness, can’t handle full dataset retrieval
- 4. VectorDB <> LLMs User Query Vectorestore LLM Retrieve similar small chunks Child Child Chunks From memory: Large chunks Context for the LLM LLM Response (max accuracy: 65%) Naive, lacks domain awareness, can’t handle full dataset retrieval ! HIGH token usage
- 5. A wide-reaching problem Published by Stanford (2024) Published by McKinsey May (2024) Published by IBM (2024) ● 65% of companies use AI, but only 15% see bottom-line impact. ● >95% of enterprises are stuck launching GenAI projects to production ● Vector-based RAG applications are only 65% accurate max Max accuracy without GraphRAG!
- 6. Microsoft Study (2024) Current RAGs are only 75% accurate max (too focused on semantic search)
- 7. LLMs Still Hallucinate Under Load If I were building a GenAI stack today, I’d start with one question: Can your retrieval system handle multi-hop logic? Most can’t. They treat retrieval as nearest-neighbor search.
- 8. 1. "Raphael, a master of the Italian High Renaissance, is known for his exquisite paintings. Though most of his works are housed in the Vatican Museums, some are also displayed at the Louvre." 2. "Leonardo da Vinci, an Italian Renaissance artist, is famous for several masterpieces, including the Mona Lisa, which is housed at the Louvre Museum in Paris. His works are among the most visited and well-known in the world." 3. "The Louvre Museum is home to several Italian Renaissance paintings, notably from artists such as Raphael and Michelangelo, reflecting the rich artistic history of Italy." 4. "Donatello, a master sculptor from the early Italian Renaissance, is renowned for his bronze and marble sculptures rather than paintings. While his influence is widely acknowledged in major museums, including the Louvre" Which Italian artist has the largest number of paintings on display currently in the Louvre? Query Vector search is inaccurate.
- 9. So, what can be done?
- 10. Optimal LLM <> Knowledge graph synergy TEXT INPUT LLMs OUTPUT KNOWLEDGE GRAPHS ✨KG-enhanced LLMs ● Structural Facts ● Domain-specific Knowledge ● Symbolic-reasoning TASKS KGs OUTPUT LLMs ✨LLM-augmented KGs ● General Knowledge ● Language Processing ● Generalizability
- 11. GraphRAG is the optimal solution.
- 12. GraphRAG is the optimal solution. Which Italian artist has the largest number of paintings on display currently in the Louvre? Query KG Output Vector Output Semantic search-based answer which offers only a partial view of the knowledge Relationship- based answer which aggregates many knowledge points
- 13. GraphRAG builds a knowledge graph from source documents Deeper understanding of the data Higher accuracy VectorRAG won’t cut it.
- 14. How it (actually) works
- 15. LLM-ready architecture Data ingestion LLM used for Entity and Relationship Extraction (ERE) I LOVE GRAPHRAG Knowledge Graph stored in a Graph Database
- 16. Synergy slide Easier to create knowledge graphs with LLMs LLM used for Entity and Relationship Extraction (ERE) I LOVE GRAPHRAG Knowledge Graph stored in a Graph Database
- 17. LLM-ready architecture Knowledge retrieval Query Knowledge Context Enterprise Knowledge Retrieval: extracting the relevant sub-knowledge
- 18. LLM-ready architecture Knowledge retrieval Query Knowledge Retrieval Enterprise Knowledge Retrieval: extracting the relevant sub-knowledge Personalized Memory Retrieval: extracting the relevant sub-long memory Memory Retrieval
- 19. LLM-ready architecture Answer Generation Accurate response with improved personalization and context-awareness Query Knowledge Retrieval Memory Retrieval
- 20. LLM-ready architecture Commit to memory Performance and Scalability Complex Query Support Integration Reduces query times Maintain context Data consistency Multi-tenancy (10K+) Designed for Privacy & Security
- 21. Policy analysis without GraphRAG is like proposing without a ring and expecting a yes! Don’t be this guy.
- 22. Who is using it, and for what?
- 23. ● GenAI developers building high-accuracy B2B apps ● Enterprises that need scalable, low-latency graph databases for mission-critical applications ● Developers building multi-tenant SaaS Who is this suitable for? See Use Cases
- 24. Fraud detection Reveal hidden patterns in your data to detect complex scams ● Detect fraud rings by analyzing relationships between entities such as IPs, devices, and transactions. ● Use real-time graph analytics to uncover anomalies, track patterns across accounts, and adapt dynamically to evolving fraudulent behaviors.
- 25. Network & Asset Relationships as a Graph ● Conventional tools store data as logs or tables—great for rows, but poor at relationships. ● Cyber threats move across systems: credentials → devices → data → users ● Graphs capture those relationships natively: ○ Nodes = users, devices, IPs, keys, services, files, containers ○ Edges = “accesses,” “belongs to,” “connected to,” “hosts,” “scanned by,” etc.
- 26. Financial data & policy analysis Unify internal silos and external feeds into actionable knowledge ● Financial policies don’t operate in isolation—they interact with markets, institutions, geopolitical events, and regulatory changes. ● GraphRAG maps complex relationships across policies, and institutions, simulating policy impacts with high accuracy. Use Case Example 1 How interest rate changes impact liquidity & compliance. Use Case Example 2 How liquidity ratios shift under new monetary policies & transaction flows. Use Case Example 3 How trade tariffs influence currency fluctuations & hedging strategies.
- 27. Stop choosing between scale, latency, and accuracy. FINANCE Fraud Detection Access Identity Management FS Data Analysis INSURANCE Policy Help Desk Claim Processing Insurance Policy Analysis RESOURCE MGMT. Asset Management Network optimization CyberSec Graph database: Use cases
- 28. Visualize codebases as knowledge graphs to analyze dependencies, detect bottlenecks, and optimize projects. ↗ code-graph.falkordb.com
- 29. Who are we?
- 30. A market-proven technology Bigger, faster, better than the competition ● Used by tier-1 enterprises ● >9 years in the market ● Benchmarks say it all: ○ 496x faster ○ 6x more memory efficient ○ 11x higher throughput
- 31. (1) AUDIO INPUT VectorRAG is killing me… TEMPORAL LOBE Audio → Text FRONTAL LOBE (GenAI) Text → “Tokens” (3) KG-RAG Text → Query Hippocampus = Knowledge Graph (4) RETRIEVAL Neocortex = Memory (5) DATA (6) RESULT SET Text → “Tokens” (7) AUDIO OUTPUT I should really try GraphRAG… 1 2 2 3 3 4 4 5 6 6 7
- 32. Thank you! ● @falkordb (LinkedIn/X) ● Connect on Discord ● 🔗 falkordb.com Visit FalkorDB