# Wisdom-Driven Knowledge Augmented Generation at Scale

Chin Keong Lam, Patho AI | AI Engineer World's Fair 2025 | 18:43

Source: https://www.youtube.com/watch?v=9AQOvT8LnMI
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/wisdom-driven-knowledge-augmented-generation-at-scale
Published: 2025-08-22
Tags: agents, multi-agent, rag, structured-outputs

## TL;DR
- Knowledge-Augmented Generation combines language models with a structured knowledge graph so an AI system can reason over relationships and give advice.
- A wisdom graph connects knowledge, experience, insight, decisions, and situations through feedback, allowing the system to improve its guidance over time.
- Knowledge graphs are better suited than vector RAG for competitive analysis and numerical questions that require multi-hop queries, structured evidence, and calculations.

## Summary
Chin Keong Lam describes how Patho AI builds expert AI systems that go beyond retrieving passages from a database. His approach, called Knowledge-Augmented Generation, adds a structured knowledge graph to a language model. The graph captures domain relationships and maps knowledge, experience, insight, decisions, and real-world situations into a feedback loop. Lam applies this model to a competitive-analysis chatbot that advises a marketing team on questions such as how to beat a competitor given current market share. He explains an implementation using n8n and a supervisory wisdom agent that coordinates specialist agents and updates a central graph. The talk also makes a practical case for graph-based systems when answers depend on numerical reasoning, multi-hop relationships, and evidence from structured data. Lam recommends a hybrid extraction process, where an LLM creates the initial graph and a domain expert prunes its taxonomy and relationships. He reports benchmark results for accuracy, flexibility, reproducibility, traceability, and scalability, with accuracy reaching 91% in his test.

## Key ideas
### A knowledge graph stores domain reasoning as relationships
[01:09](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=69s)
Lam defines a knowledge graph as a way to preserve wisdom by connecting concepts into a network. The graph captures the thought process and taxonomy of a specific area of expertise. That structure matters when an AI system needs to think and provide advice instead of simply retrieving data. In his framing, a language model becomes more useful when it can work with the relationships between entities, decisions, situations, and domain concepts. The graph is intended to encode how experts understand a problem, rather than only storing the facts that an expert might look up.

### Wisdom guides decisions by interpreting real-world situations
[02:32](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=152s)
Lam's diagram puts wisdom at the center of a loop. Wisdom guides decision making, while decision making examines a situation in the real world. Knowledge feeds wisdom, and wisdom synthesizes information after it has been ingested. He gives social media sentiment as an example of a chaotic input from which the system can derive patterns about products and competitors. In his model, knowledge tells you what something is, experience tells you what worked before, and insight suggests what to try next. Feedback from situations, experience, and insight then changes the wisdom used for later decisions.

### The competitive-analysis chatbot maps business concepts onto the wisdom loop
[06:26](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=386s)
Patho AI applied the model to a client's competitive-analysis work, which had previously been done by a marketing department. The chatbot is designed to answer strategic questions such as how to win against a competitor in a market. Lam maps market data to knowledge, past campaigns to experience, industrial insights to insight, and the company's current sales situation to the situation node. Competitor weaknesses also inform the situation. A wisdom engine orchestrates these elements and advises the system on what to do next, instead of treating the chatbot as a simple question-answering interface.

### A supervisory agent coordinates specialist agents around a central graph
[08:37](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=517s)
Lam uses n8n to prototype the system as a workflow made from Node.js components. An AI agent acts as the wisdom agent and can drive models from OpenAI, Anthropic, and local deployments. It supervises other agents that handle particular parts of the state diagram. An insight agent, for example, examines social media for sentiment about products and sends its findings to a central graph. Multiple agents update their own perspective in the shared graph, while the taxonomy gives the system a common structure for making decisions. Lam says the graph structure and taxonomy matter as much as the model used for the final response.

### Knowledge graphs fit competitive analysis because they preserve relationships
[11:37](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=697s)
Lam gives several reasons to use a knowledge graph for competitive analysis. Graphs capture complex relationships between entities, which provides more context about gaps and competitor advantages. Structured data and semantic relationships can reduce noise and improve the precision of generated answers. Graphs can also absorb new data sources and relationships as the taxonomy develops. Their query systems can traverse several relationships for multi-hop questions. Lam also says a graph can integrate text, images, and video when the data is placed in the right structure, giving the system a broader view of the material used for strategic decisions.

### Vector RAG struggles when the answer requires numerical calculation
[14:01](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=841s)
Lam says vector RAG is good at semantic similarity but performs poorly on some numerical reasoning tasks. If asked for Apple's revenue across two years, a vector store may return passages about the company instead of calculating the requested change. In his example, a structured knowledge graph contains the relevant financial figures. A query engine selects the 2021 and 2022 revenue values and calls a function to calculate the result, producing 15.23. The advantage is evidence-based reasoning from structured figures rather than an answer assembled from a loosely related retrieved passage.

### A hybrid system can route simple questions to RAG and harder ones to graphs
[15:34](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=934s)
Lam does not present the graph as necessary for every use case. A simple product-information question can use a Chroma database with an LLM agent. A question about beating the competition given the company's current market share requires a graph database, Cypher queries, and several rounds of multi-hop retrieval. His proposed architecture combines regular RAG with a wisdom graph and chooses the heavier path when the question needs relationships, calculations, or strategic reasoning. This lets the system match its infrastructure to the question rather than applying one retrieval method to every request.

### Experts should prune LLM-extracted graphs before relying on them
[16:45](https://www.youtube.com/watch?v=9AQOvT8LnMI&t=1005s)
Lam compares fully automated graph extraction with manual construction and recommends a hybrid approach. An LLM can first extract the graph, but the domain expert who defines the taxonomy should review it and prune unnecessary relationships. Pruning removes connections that do not belong in the model's reasoning structure. He reports benchmark scores for the resulting system: 91% for accuracy, 85% for flexibility, followed by measures for reproducibility, traceability, and scalability. His conclusion is that graph quality depends on the framing and taxonomy used to store wisdom, not only on the extraction model.

## Notable quotes
- "KAG doesn't just retrieve, it understand." (02:07)
- "Knowledge tells you what it is, experience tell you what worked before. Insight invent what to try next." (04:30)
- "Vector store excel at semantic similarity but struggle with complex numerical calculation." (14:15)
- "The most important part of the knowledge graph is feedback loop." (04:58)
- "After you use the LLM to extract your graph you ask the expert that you're going to build to build your taxonomy to prune the graph." (17:05)

## Tools & references mentioned
- Patho AI
- National Science Foundation
- SBIR
- n8n
- Node.js
- OpenAI
- Anthropic
- LangChain
- Chroma
- Neo4j
- Cypher
- LLM Graph Transformer

## Who should watch
- You are building an expert or advisory system and need it to reason over domain relationships instead of returning retrieved text.
- Your application answers competitive, financial, or other numerical questions where structured evidence and function calls matter.
- You are deciding whether to add a knowledge graph to a RAG stack and want a practical prototype architecture using agents and n8n.

## Related talks

- [On AI and Knowledge](https://aietalks.com/talks/on-ai-and-knowledge) (Pablo Castro, Microsoft, 17:35)
- [Intro to GraphRAG](https://aietalks.com/talks/intro-to-graphrag) (Zach Blumenfeld, Neo4j, 1:18:35)
- [Anchoring Enterprise GenAI with Knowledge Graphs](https://aietalks.com/talks/anchoring-enterprise-genai-with-knowledge-graphs) (Jonathan Lowe, Pfizer & Stephen Chin, Neo4j, 20:59)
- [Practical GraphRAG: Making LLMs Smarter with Knowledge Graphs](https://aietalks.com/talks/practical-graphrag-making-llms-smarter-with-knowledge-graphs) (Michael Hunger, Jesús Barrasa & Stephen Chin, Neo4j, 19:46)
- [GraphRAG: The Marriage of Knowledge Graphs and RAG](https://aietalks.com/talks/graphrag-the-marriage-of-knowledge-graphs-and-rag) (Emil Eifrem, Neo4j, 19:15)