↔ Drag to browse
No videos.
Pentum_Graph explores how unstructured information can be automatically acquired, connected within a graph, and transformed into a persistent information space that can be queried using natural language.
Pentum_Graph explores how unstructured information can be automatically acquired, connected within a graph, and transformed into a persistent information space that can be queried using natural language.
The initial research focus is publicly available YouTube videos. Videos are a complex form of unstructured data, combining visual content, audio, language, and temporal information.
Videos are processed automatically and represented as nodes, while detected relationships are represented as edges. This creates an information structure that exists independently of the original platform.
Graph → Relationships, context, and clusters
Timeline → Information within a specific context displayed over time
ExoSeek → Graph-based search through connection paths
Sources → Information sources and their relationships to the underlying information
The research is not limited to YouTube. The long-term goal is to integrate information from different sources and data formats into a single graph while preserving provenance and traceability.
The broader vision is a personal information space for each user.
For example, if a user specifies “Baseball” as an area of interest, the system first identifies relevant people, organizations, competitions, events, and relationships. From this, it builds a topic map and an initial information state.
From that point onward, the system continuously performs the following process:
Search → Acquire → Analyze → Classify → Connect → Evaluate → Update → Deliver
New information is not simply stored. It is evaluated in relation to the existing information state.
The graph becomes a persistent information model that preserves information, relationships, sources, context, and changes over time.
The LLM serves as the natural-language interface to this graph. Rather than continuously “training” the model itself on newly acquired information, the system dynamically retrieves the relevant context from the graph in response to each query.
Autonomous system = Information acquisition and updating
Graph = Memory, structure, and context
LLM = Interpretation, reasoning, and interaction
The central research question is:
Can continuously acquired, unstructured information be transformed into an automatically updated personal information state that can be queried using natural language?
A user begins by telling Pentum_Graph that they are interested in “Baseball.”
From there, the system investigates the major entities and relationships associated with the topic, including players, teams, leagues, competitions, games, trades, injuries, and statistics. It then builds an initial structure for the topic.
Interest → Topic Map → Baseline Information → Continuous Observation → New Information → Connection to Existing Information → Graph Update
For example, if a player moves to another team, that information is not stored merely as an isolated article or event.
Instead, it is connected to the player, the former team, the new team, the league, the contract, related people, previous events, the information sources, and the time at which the event occurred. The existing information state is then updated accordingly.
If multiple sources report the same event, those sources can also be linked to the same underlying context.
Over time, the result is not simply a collection of information. It becomes a continuously evolving information state representing what is known about baseball.
The user can then query that state using natural language, for example:
The goal, therefore, is not to build another search engine or content feed.
The goal is to continuously observe an area of interest, structure and connect the information that emerges, keep that information state up to date, and make the accumulated state itself queryable.
That is the core idea being explored through Pentum_Graph.
No videos.