Tutorial: Building with CleaveDB

Semantic Pathfinding

Traditional graph databases traverse paths purely symbolically—matching exact edge labels with no semantic understanding of the data inside nodes. Vector databases search semantic meaning, but lack structural relationships. CleaveDB pioneers Neuro-Symbolic Semantic Pathfinding by blending both paradigms into a single query execution engine.

Concept-guided graph traversal

Use GUIDED BY MEANING to steer a traversal toward documents with relevant semantic context:

FOLLOW "users:alice" THROUGH "interests"
  GUIDED BY MEANING "cloud infrastructure"
  LIMIT 5

Rather than returning every linked interest uniformly, CleaveDB evaluates the neural cosine similarity of each adjacent node against the concept "cloud infrastructure" and prioritizes the most conceptually relevant paths.

Hardware-accelerated SIMD pruning

You can attach confidence cutoffs using THRESHOLD to prune non-relevant branches early:

TRACE "projects", "topics" FROM "users:alice"
  GUIDED BY "machine learning"
  THRESHOLD 0.75

At each hop, CleaveDB’s Rust engine executes the C++ _mm512_dp_ps AVX-512 SIMD vector dot-product in hardware. Branches with cosine similarity below 0.75 are discarded in a single clock cycle, eliminating graph explosions before they occur.