RAG over a graph — in one Cypher query

xrayGraphDB ships native embed() and vector_dot_product() as first-class Cypher functions. Type a free-form symptom below — the engine embeds your query, runs vector similarity against 274,065 real FDA warning labels, then graph-traverses (:Drug)-[:HAS_WARNING]->(:Warning) to surface every drug whose warning matches. All in one transaction. One query. One engine.

How it works — the exact Cypher

Two documented Cypher functions do all the heavy lifting. Both are visible at docs.html#fn-rag_llm — nothing proprietary, nothing hidden.

Step 1 — Embed once (offline)

For every :Warning node, store a 384-dim vector generated by the engine's native ONNX-or-hash embedder. No external API call. Run once at ingest time:

MATCH (w:Warning) WHERE w.embedding IS NULL
WITH w LIMIT 1000
SET w.embedding = embed(w.text)
RETURN count(w) AS embedded;

Step 2 — Search at query time

Embed the user's free-form text inline, dot-product against every warning's stored embedding, traverse the graph to the linked drugs, return the top matches. One round-trip. One transaction. No two-system sync.

WITH embed($user_query) AS qv
MATCH (d:Drug)-[:HAS_WARNING]->(w:Warning)
WHERE w.embedding IS NOT NULL
WITH d, w, vector_dot_product(qv, w.embedding) AS similarity
ORDER BY similarity DESC LIMIT 10
RETURN d.name, w.text, similarity;

That is the entire RAG pipeline. The query the audience copy-pastes is the query that ran above.

Why one engine matters

The full RAG/LLM toolkit (29 functions)

This demo uses two. The rest are documented at docs.html → RAG/LLM and run identically — native Cypher, no extensions:

Data is real, not seeded

Every Warning shown is a real FDA label from open.fda.gov, ingested via the same public pipeline at demos/pharma/ingest/. No mock data, no curated subset to make the demo look good. The query you run hits the same data every other pharma demo query hits.