Knowledge graphs in 5 minutes

A knowledge graph is just two things: dots and lines.

• The dots are called entities: people, organisations, drugs, laws, concepts. Any "important noun" in your corpus.
• The lines are called relationships: how those entities connect. "Acme picked Postgres", "the law regulates treatment X", "Alice works with Bob".

That's it. What GRAIL does is build that graph automatically from your text.

Think of GRAIL as…

Imagine that instead of an alphabetical list of every book in a library, you have a wall map where each author is a dot and lines connect authors who cite each other, books that share a topic, characters that appear in multiple works.

For questions like "which books talk about Don Quixote?" the alphabetical list is barely useful. The map answers by walking the nodes instead of reading pages.

Why it matters for Q&A

Traditional RAG says: "I'll search for pages most similar to your question and hand them to the LLM". That works for questions whose answer fits in one text chunk.

But it fails when:

• The answer crosses documents ("how does X from this report relate to Y in that one?").
• The question is structural ("which drugs does oncologist Pérez's protocol cover?" — you have to walk oncologist → protocols → drugs).
• You need thematic synthesis ("what's this whole thing about?").

A graph makes those questions structural, not probabilistic. If Acme is connected to Postgres with a "CHOSE" edge, the answer to "which DB did Acme pick?" is a one-hop walk, not a similarity search.

What GRAIL adds on top of the basic graph

A "bare" graph (entities + relationships) is already useful. GRAIL layers three more things on it:

1. Communities

GRAIL runs the Leiden algorithm to group densely connected entities into communities. It's like dividing the library into themed sections, automatically and at multiple granularities.

2. Community reports

For each community, an LLM writes a narrative summary of what it's about. It's the "section pamphlet" the librarian hands you when you ask something broad.

This is what makes global mode work: the answer to "what are the central themes?" isn't reading the whole corpus — it's reading the reports.

3. Retrieval queries on entities

Each entity stores 2-3 anticipated questions in its embedding text. It's the "post-it the author stuck on each book" saying "read me if you care about X".

This dramatically improves cross-lingual and intent-based matching. A vague question matches well if the right entity has a post-it that resembles it.

Why GRAIL is more expensive up front (and cheaper in use)

Building the graph costs one LLM call per chunk of corpus. For a corpus of 200 PDFs, that's hundreds or thousands of calls during indexing.

But after that, in use, each question costs one call (or a few with agent). Vanilla RAG also costs one call per question — but answers structural questions worse.

The math tips in GRAIL's favour when:

• You're going to ask many questions of the same corpus.
• The questions are structural or synthesis.
• You care about quality more than per-question cost.

Next step

• Search modes — how the graph translates into six answer strategies.
• Cascade — the mode that combines graph with text for factual questions.
• Communities and Leiden — more detail on how communities form.