Advanced Context Engineering with Knowledge Graphs
As we all know, the answer of an LLM is as good as the context it is provided with. If the information is not present in the context, the LLM is likely to hallucinate an answer which is that more dangerous that it is plausible. With the recent increase to the context windows, it might be tempting to solve this issue by dumping large amounts of information into the context. But this is ill-advised for two reasons:
- Billing will raise because of the increase of input tokens.
- The LLM will oftentimes get confused if the relevant information is drowned in noise.
Hence, context engineering should be considered the bedrock of reliable LLM-based applications. The most well-known approach consists of maintaining a flat list of information chunks. This list is then embedded into vectors to perform semantic matching for context retrieval. The topic of this webinar is how to leverage a more structured way of representing information for context retrieval.
We believe that knowledge graphs offer a valuable approach in this respect. A knowledge graph is a structured representation of information where entities (such as people, places, or concepts) are connected by relationships, forming a graph that encodes meaning. For instance:
(Marie Curie) -- won -- (Nobel Prize)
The webinar will feature three presentations and include time for Q&A:
- Laurent Christophe (Sirris) will give a brief overview of what knowledge graphs are, how they can be populated and how they can be queried.
- Aziz Nebli (Sirris) will make a brief deep dive into a real world example of knowledge graph used for a conversational interface.
- Semantic Context for Agents - Carlos Noguera (Logic.Tools): In this talk, Carlos will introduce OWL Ontologies as a way to describe a knowledge graph. Ontologies are an expressive language with which to model the information contained in nodes and edges of a KG. Furthermore, Carlos will demo how ontology-augmented KGs, when coupled with a reasoning engine like Logic.Tools's, provide rich, semantic context to conversational agents.