Coreon’s Post

If you are not using Knowledge graphs with AI agents, You're missing out, Here's an example from Amazon... Unlike vector databases, Knowledge Graphs (KG) are best for connecting different data points with each other. This allows for efficient retrieval of information from the documents. Let me show you an example with KGLA (Knowledge Graph Language Agents) 📌 Here's how it works: 1. Path representation: - Treating users and items as nodes in KG. - and the relationship between these nodes are represented as paths within the Graphs. 2. Agentic Simulation: - It then employs multiple language agents that simulate user and item interactions. - Each agent maintains a memory that records the profiles of users or items. - During the simulation, the agents use KG paths as simple language to interact and understand the reasons for their choices. Example: If a user states, I like monsoon apples. It identifies the knowledge path user->apple->green->monsoon->sweet to understand that users like green monsoon apples because they are sweet in taste. 3. Path translation and incorporation: - After the modules have understood the path behind the user's query. - They extract the path which is then translated for Language agents to understand. - Then they employ another module that integrates the translated paths into the agents' decision-making processes.   4. Simulating the user thinking and response: - Now the agents simulate a path of how user would think before giving a response with KG. - Now from that path, they would incorporate necessary elements required for more comprehensive answer. - After than, they just output the required answer. 🔢 Now lets, see how useful this process is: 1. Unlike Agent CF, KGLA with 2-hop + 3-hop KG Paths outperforms with nearly 2.3 times. - Baseline Agent CF on NDGC@1 = 193 - KGLA with 2-hop + 3-gop KG path on NDGC@1 = 0.377 2. Similar to that, KGLA even outperforms BM25 - Baseline BM25 on NDCG@5 = 0.362 - KGLA with 2-hop + 3-gop KG path on NDCG@5 = 0.637(76% improvement) 💭 Personal Thoughts: 1. As I said, Utilizing Knowledge graphs will become the next trend for AI Agents. 2. Even in this use case, a proper path is always useful for proper reasoning behind the answer generation. What do you think about this new dataset and the approach? Let me know in the comments below 👇 Please make sure to, ♻️ Share 👍 React 💭 Comment to help more people learn P.S. Check the comment for references