InfraNodus is effectively a portable GraphRAG system that enables you to use a knowledge graph to explore and develop new ideas. Instead of using a standard LLM RAG system that's based on finding similar statements to your query, GraphRAG has a relational understanding of the context through an underlying knowledge graph that can be used to get a general, "holistic" insight about the underlying data and retrieve pertinent relations from text.
As an InfraNodus user, you can utilize the graph interface to explore the topics and concepts that are interesting to you and to direct LLM to think in a specific direction of your choice.
What is GraphRAG, and how is it different from RAG?
Normally, when an LLM is asked to provide a response, it's going to ingest the context and perform a similarity search, finding the statements that are semantically close to the original prompt and using them to generate the final response. This approach is called RAG (retrieval-augmented generation) and has multiple problems, specifically when the user queries are too general or require the model to retrieve contextual relations from the original text. For instance, if you ask "What is it about" to a knowledge base that contains this support portal, the retrieved chunks will be semantically similar to "what" and "about" but the base model will actually not understand that the user is inquiring about the capabilities of the tool.
This is where Graph RAG comes in. It helps the model to have a general "holistic" understanding of the context and to extract additional relationships from the text that represent.
When a user makes a query, it's not sent to an LLM directly. Instead, it's
- converted into a graph,
- then it's compared to the underlying knowledge graph structure that it is enquiring,
- then a part of the underlying graph that overlaps with the user query graph is extracted, along with 1 or more hops along the edges, containing the statements that describe those relationships
- also general information about the context is extracted, such as the main concepts and topics inside
- this rich context is then used to augment the original prompt and to help the model to have a better understanding of the context
In the end, what the model "sees" looks something like this:
In this example, if a user is asking about "generating ideas" with InfraNodus, the system is finding an overlap with the graph, sees that there is more relevant content in relation to "AI", "analysis", "content", "network", "topic", then extracts relevant relations and statements as well as the general topical structure.
The user's prompt is augmented with this data, so that when the actual answer is generated, it will use this additional rich context and not just the statements about "generating ideas" from the original text.
As you can see, unlike other GraphRAG / HyperRAG implementations that can only be used by LLMs, InfraNodus can also be used by humans. The interactive knowledge graph is fully interactive: you can select the relevant topics, concepts, and gaps, focus on specific relations and have the built-in LLM model generate insights for you. It is also possible to feed this data to an LLM workflow using the InfraNodus API and to embed these knowledge graph insights into other AI-based workflows.
You can learn more about how the InfraNodus GraphRAG implementation works in the Nodus Labs article.
How to Use the InfraNodus Graph to Steer LLM Thinking Process
Here are some of the ways you can apply InfraNodus GraphRAG for practical use cases.
1. Improving a general understanding of the context and developing it further
When you're looking at a particular context or asking a question about a text, both you and your LLM will benefit from having a general understanding of the main topics inside. The main topical clusters identified by InfraNodus are used by its API to augment the prompts, so when you ask a question or generate a summary using the built-in AI, this additional data is used to improve the results. As a user, you can also use the visualization to improve your understanding of the context and knowledge base observability:
In the InfraNodus API, this corresponds to the general (default) mode `optimize=develop` where all the main clusters of the graph are used to provide the best possible response to the query.
If you choose the `optimize=reinforce` mode, then AI will focus on the top 4 clusters and on the least connected cluster, attempting to optimize the structure.
The response provided by the API in this case will be focused on the main topical clusters as specified above.
2. Ability to focus on specific topics and develop ideas contained in them
You can use the graph to focus on a specific concept and develop ideas inside. In order to do that, you can select the topic in the graph itself (or using the API) and InfraNodus' API will focus on that particular topic or idea:
In the InfraNodus API, this mode can be initiated when you provide a list of the topics (as an array of IDs from `1` to `7`) or words (as an array of keywords) that should be selected in the graph. Then the API response will be focused on those particular topics / ideas.
Using the `optimize=gap` mode, one can choose the topics automatically based on the structural gap in the network.
3. Ability to find the gaps between the topics to generate new ideas
You can switch to the "gaps" view to uncover the structural gaps in the knowledge graph (clusters of ideas that are not very well connected in this particular context). This is akin to selecting the topics in the section 2, except that in this case InfraNodus will automatically select the topics that are not well connected. You can then use AI to develop this content further by auto-generating question / idea / summary or switching into AI chat mode to interact with the graph but focusing on these particular ideas.
In the InfraNodus API, this corresponds to the `optimize=gaps` mode, which retrieves the gaps automatically. You can also specify the gap with the `gapDepth` parameter (from 0 to 7, which corresponds to the Gaps 1 - 8). The API will provide a response that will focus on the gap.
Conclusion
As you can see, the knowledge graph can be used to steer the model towards specific topics, ideas, and gaps in the underlying structure. It can be a very poweful tool to explore any knowledge base or to brainstorm on a certain topic using a specific ontology as a navigational device for thinking.
We can especially recommend using the gaps feature and selecting the nodes that are not yet connected in order to generate new ideas and insight.
Comments
0 comments
Please sign in to leave a comment.