Robert Caulk argues that unconstrained knowledge graphs can capture high-resolution relationships in news without forcing every possible relationship into a fixed ontology.
2
AskNews builds synthetic representations of news articles with a pipeline that combines Llama 3.1 70B, GLiNER, a fine-tuned 3B model, keyword search, vector representations, and Memgraph.
3
Diverse sources and competing viewpoints can help users and language models identify contradictions, while human review and reporting-voice filters provide additional checks against misinformation.
Summary
Robert Caulk explains how Emergent Methods builds AskNews, a real-time news retrieval service based on a large, unconstrained knowledge graph. Instead of forcing every article into a fixed ontology, the system extracts people, places, evidence, metadata, and relationships at high resolution. It stores much of the raw information in S3, retrieves relevant material with conventional search and vectors, then creates small, temporary Memgraph instances for particular user questions. The system uses Llama 3.1 70B, GLiNER, and a fine-tuned 3B model to construct synthetic article representations. Caulk argues that this structure is useful for giving language models compact, grounded context and for comparing competing accounts of the same event. He is open about the tradeoffs: ontology-free graphs are harder to query and can become slower as they grow. The conversation also covers source quality, human editorial review, reporting-style filters, publisher compensation, and the limits of news data for events that unfold first on social media.
Ontology-free graphs capture relationships that fixed schemas cannot anticipate
Robert Caulk says Emergent Methods chose an ontology-free approach because news contains an effectively unlimited range of relationships. A traditional graph might encode that Donald Trump appears in an article, that the article has a publisher, and that it was published on a date. It would need a new predefined relationship for claims such as "Trump defeated Kamala" or for links between people, organizations, and political groups. An unconstrained graph lets the relationships form from the source text. Caulk says this produces a higher-resolution representation that can be passed to an LLM as compact JSON instead of a long paragraph. The tradeoff is less predictable querying and more difficult traversal.
A hierarchy of storage systems keeps large graphs usable
Caulk says an ontology-free graph cannot simply be placed in a conventional graph database with an unlimited number of indexed relationships. AskNews stores high-fidelity information across a hierarchy of systems, with much of the data in S3. Traditional indexing, metadata filtering, keyword search, and vector representations retrieve a smaller relevant section. That section can then be placed into a graph database for traversal. He describes Memgraph as useful because it is open source, fast, and designed around in-memory performance. The goal is to work with the right graph for a question instead of presenting users with millions of irrelevant relationships.
Small temporary graphs are created for each user investigation
When a user submits a natural-language question, AskNews identifies useful keywords and concepts, retrieves the relevant information, and builds a small Memgraph instance. A question about the people connected to Biden within three hops might require retrieval, graph construction, and a query generated from the user's English request. The user can receive a summary or access the graph to run additional queries. Caulk says these graphs may cover only recent events, such as the previous 72 hours or week, because users usually need a focused and timely view. Once the investigation is finished, AskNews deletes the instance, although it keeps limited caching.
Article processing creates a synthetic representation before graph retrieval
AskNews processes every article by extracting evidence, important people, geographic coordinates, and relationships. The system must also handle different references to the same person, such as "Joseph R Biden," "President Biden," and "Biden." Caulk describes a pipeline that uses Llama 3.1 70B for a broad extraction and enrichment step, GLiNER for entity extraction, and a fine-tuned 3B model whose only job is generating an ontology-free graph. The data is stored with identifiers across several systems. Keyword search remains useful for initial retrieval, while later disambiguation and graph traversal compress the information into a more usable structure.
A focused small model makes high-volume graph construction affordable
Caulk says the team used GPT-4o to label a diverse training set, covering politics, sports, entertainment, languages, continents, countries, and different source types. They then fine-tuned the 3B model on that data. He says the resulting model outperforms Claude 3.5 on the graph-building task while running quickly on a consumer graphics card. The specialization matters because AskNews processes roughly half a million articles per day, according to Caulk's cost comparison. Sending that volume through a larger general-purpose model would be too expensive. The small model does not perform every task, but it handles graph construction well.
Compact graph context lets downstream LLMs use more of their context window
Many AskNews users do not need the temporary graph database. They request a relationship graph, receive it as JSON or YAML, and place it into their own LLM prompt. Caulk gives the example of building relationships between Biden and Hunter Biden from thousands of articles, then reducing the result to a few hundred tokens. The customer's application can add private context and ask another model to produce an answer. AskNews handles the aggregation and real-time grounding without sending the original articles into the downstream prompt. Caulk also says the synthetic representation gives publishers a way to receive royalties without exposing the original journalistic expression for copying or model training.
Competing sources make contradictions visible instead of hiding disagreement
Caulk says AskNews does not need to remove every source that readers distrust. For a topic such as the Russia-Ukraine war, the system can include Russian, Ukrainian, French, German, English, Italian, and Arabic sources, then show where their accounts align or conflict. A statement from RT can appear alongside other reporting, with the contradiction made explicit. Users can filter sources by reporting voice, including objective, persuasive, sensational, investigative, or analytical. Caulk says Llama 3.1 70B performs this enrichment and sentiment analysis, although he does not present it as perfect. Human review also covers sources, data, and output.
News graphs help investigations by exposing indirect connections
Caulk describes clients using the graph to follow connections that ordinary vector search might miss. Love Justice, which works against human trafficking in South Africa, can begin with a reported criminal incident and follow links among people, shops, locations, and other events to find an unusual connection worth investigating. He also mentions graph-based criminal investigations involving Venmo transactions, open-source intelligence, imagery, ports, boat registrations, and Bitcoin wallet movements. In these cases, the value comes from following a path through related entities and transactions. The connection may be weak or unexpected, but it can give an analyst a lead that was not visible in an isolated document search.
The ontology-free choice gives flexibility at the cost of direct querying
When Demetrios Brinkmann asks for the downside, Caulk says ontology-free graphs become harder to handle as more data is added. Traversals can become slower, and query construction is less direct because there are no predefined relationship definitions to guide propagation. A structured ontology makes it easier to reach a known part of a graph and supports creative graph operations such as link prediction. Caulk's view is that ontology-free representations are especially useful when the goal is to communicate highly resolved information to an LLM. The approach sacrifices some query efficiency and predictability for broader coverage of relationships that were not known in advance.
"The strength is with this high resolution, I can now look at a very small piece of my graph and quickly consume that information as an analyst or pass that to an LLM in a very, very prompt-token-optimized way."Robert Caulk06:33
Who should watch
You are building retrieval-augmented generation systems and need to decide whether a graph database is justified for your workload.
Your application needs to compare news sources, expose conflicting accounts, or follow indirect links across people, organizations, places, and events.
You are operating a news or research product and want a concrete discussion of source filtering, publisher compensation, model costs, and the limits of real-time coverage.
