A political knowledge
graph, built in a week.

A body of policy
too big to hold.

A policy lead inside a UK political party had built a deep, fast-growing body of work, positions, research, transcripts and documents spread across drives and people's heads. Getting a consistent, sourced answer to "what's our position on this, and why" meant digging every time, with the constant risk of contradicting something said before. The thinking was all there. It just wasn't anywhere a tool, or a teammate, could reach.

One graph,
in five days.

Using a 5-day framework, we built the policy lead a knowledge graph they own. We pulled from fifteen distinct categories of political data, public statements, policy documents, transcripts, social data and more, structured everything with their own five-layer policy framework, then ingested it into a graph they can interrogate by simply chatting with an LLM. It answers from their data rather than intuition, returns sourced, well-written outcomes, and even surfaces the hidden geometry of how positions relate.

The five-layer framework is what turns raw political data into structured ideological intelligence. The schema uses edges to capture affiliations and positions, so the graph can map content across ideological axes and support real qualitative analysis, not just keyword lookup. Every answer carries the supporting body of knowledge that feeds it, which is why it holds a consistent line instead of contradicting past statements.

The policy lead led the build; I mentored. So it's not a black box they rent, it's a system they understand and extend themselves, every week.

First, the methodology, the five-day path we followed from scattered sources to a working, owned graph:

flowchart TD
  A["Day 1: raw material, transcripts, documents, policy papers"] --> B["Day 2: structure with the five-layer policy framework"]
  B --> C["Day 3: ingest into a knowledge graph"]
  C --> D["Day 4: query and validate, grounded, sourced answers"]
  D --> E["Day 5: handover, she owns and extends it"]

How the live
solution runs.

The five-day framework above is how we built it. The architecture below is a different thing: it's how the finished system actually runs, day to day.

We pulled 25 accounts across YouTube and TikTok with Scrape Creator, capturing not just videos and transcripts but the comments too, the audience's reaction, not only the message. That fed an ontology and schema we worked hard to get right, so nothing important was lost and even the hidden values in the context, policy positions, ideological framing, were captured. We chose FalkorDB as the graph store, for speed and the built-in graph mechanics we could build on (here's how it compares to Neo4j), and built agents in VS Code that query it through the MCP connector in plain natural language, against the schema. Nothing exotic, just code that does the analysis.

flowchart TD
  A["25 accounts across YouTube + TikTok"] --> B["Scrape Creator: videos, transcripts and comments"]
  B --> C["Ontology and schema: topics, positions, ideological framing, people, dates"]
  C --> D["FalkorDB graph store"]
  D --> E["Agents in VS Code, querying via the FalkorDB MCP connector"]
  E --> F["Reports: engagement, trends, causality, consistency over time"]

Three things mattered more than anything to keep the graph trustworthy to query:

Where the energy
actually is.

Because we captured reactions as well as words, the same graph answers questions a folder of documents never could. The policy lead can ask it:

• which content is the most engaging, and the specific cases behind it;
• how engagement relates to communication tactics, as a report, not a hunch;
• the recurring patterns and topics that consistently attract an audience;
• which people drive the most engagement with that audience;
• and whether the party's position stays consistent over time, against public sentiment and the news.

Sourced, consistent,
and owned.