Why Should Gen AI Adopters Switch to Graph Based AI Agents_.pptx

How does Vector RAG fare against Graph RAG for AI accuracy and reliability? The Tars and CogniSwitch Inc. Team practitioners pitted two AI Agents against each other for complex customer support scenarios in a Show-and-Tell session.

1. Why Should Gen AI Adopters Switch to Graph Based AI Agents?

2. Who is this webinar for? Product Managers and Engineering Managers who want to use GenAI applications to compress your workflows Engineering Teams who offer GenAI Services to your clients and are struggling with hallucinations GenAI app developers

3. Speaker Intro Josh - Founder, CTO at Cogniswitch.ai Josh is a computational linguistic domain expert and has been working on AI and NLP for over 10 yrs Prior to CogniSwitch, Josh was Head of Product at Aikon Labs Joshua Thomas CTO, CogniSwitch.ai

4. Speaker Intro Vinit Agrawal is the Founder and CTO at Tars. He leads the engineering team at Tars and has spent over 9 years building Conversational AI solutions to customers across diverse industries, including Government, Finance and Banking, Healthcare, and Insurance. Under his leadership, Tars developed and deployed advanced AI agents by leveraging Generative AI to automate and augment internal and external facing customer workflows. Presently, he's focusing on addressing Generative AI challenges like hallucination, AI explainability, and consistency by leveraging the latest research in Generative AI.

5. Present State of AI Agents - Customer Concerns

6. Present State of AI Agents - Customer Concerns E-commerce Customer Story Higher-Education Customer Story

7. Show and Tell - Vector vs Graph RAG

8. So, we built two ChatBots and trained them on Software Manual of a Property Management Software - Agilysys

9. This was done as part of a demo for CIO of an International Hotel Chain. This org is in the middle of a digital transformation project where they are replacing their Property Management Software.

10. Let’s ask questions that a front desk employee of a hotel would ask when they try to use the Property Management Software

11. https:// chatbot.hellotars.com/conv/MhIYiO https://chatbot.hellotars.com/conv/v9mBWr

12. Vector RAG CogniSwitch + Tars Q1 Accurate but lacks precision and alternative methods. Precise answer that also mentions alternative methods. Q2 Isn’t able to answer the first question. Graph RAG is able to connect multiple factors and state a step-by-step process. Q3 Isn’t able to answer the question. Graph RAG is able to connect multiple factors and state a step-by-step process for a multi-hop question. Q4 The answers states some information that’s highly relevant. Graph RAG on the other hand is able to generate a precise and accurate answer. Q5 Isn’t able to answer the question. Able to connect the two factors and state a step-by-step process.

13. Show and Tell - Vector vs Graph RAG

14. How Vector RAG Works & Limitations

15. Vector Search Over Keyword Search Keywords in user queries often don’t match with the keywords in the indexed documents For Example: User Query: “ "ways to save money for the future ” A keyword-based retrieval system won’t match with "retirement accounts" or "long-term savings plans." Vector Search comes to rescue. It will match similar concepts without exact keyword match Quick and easy to get started with a ton of unstructured text data thrown at it, and it will still work much better than keyword search

16. How Vector RAG Works https://excalidraw.com/#json=UQXZ1JH_Hbpj9lDTV9Wc-,3HVzik_4CAJJqFDUI3TWpg

17. Limitations of Vector RAG - Issues & Challenges Loss of context when chunking: Data is split into smaller crude chunks for embedding, the result is loss of context. Similar but Contradictory chunks: It can provide contradictory information if the original KB has it. As both the chunks will be matched by similarity search. Lack of reasoning capabilities: Numerical similarity algorithm is used for search, no symbolic or logical reasoning is involved. Retrieval Recall: It misses relevant info from the large unstructured text corpus. Assumes single and localised answers: cannot handle ambiguous or subjective questions, that requires deeper data inspection.

18. How Vector RAG Works & Limitations

19. How Vector RAG Works & Limitations

20. How Vector RAG Works & Limitations

21. How Vector RAG Works & Limitations

22. Sensitivity to Noise and Outliers: Just adding some words in a chunk can confuse the matching algorithm. Cost of embedding and running: Often all the vectors are kept in memory when looking for similar vectors. This can be very costly in terms of infrastructure. Determining Optimal value of K: Selecting a bad value can result in overfitting or underfitting the resulting data. Finding right K value requires a lot of experimentation. Handling different data formats: Embedding the tables, numbers, etc. as text is a very crude way, instead of respecting their tabular structure. Resulting in bad knowledge representation, and bad retrieval data. More Limitations of Vector RAG

23. The level of hallucination, or the accuracy of the answers, in a language model is dependent on the quality of Retrieved data and how that data was ingested. If the data sent via vector-retrieval to the LLM is incorrect (for the reasons discussed above), then the output from the LLM will be poor. For this reason, vector-based retrieval may NOT be the best option for enterprise use, as it isn’t always able to provide reliable and accurate results. That’s when new techniques like Graph-RAG comes in. LLM Generation: Garbage in, garbage out

24. How Graph RAG Works and Can Eliminate Hallucinations Joshua Thomas CTO, CogniSwitch.ai

25. Wait, Eliminating Hallucinations is NOT just about using a Graph DB

26. What a Rich Knowledge Structure looks like?

27. Next Steps Build your first AI Agent on Tars https://bit.ly/MakeAIAgents Join our Discord Community https://bit.ly/TarsDiscord Build Reliable Agents Using CogniSwitch Platform https://www.cogniswitch.ai Explore CogniSwitch Packs on Langchain & Llamaindex