NEBULA
docs$neb request-access

NEBULA-CONCEPTS-VECTOR-INDEXING(1)

Nebula Manual

NAMEvector indexing — Nebula is designed to be vector-native, but the current implementation is a path/hash index preview.
SOURCEnebula/docs/vector-indexing.md
STATUSpre-alpha · protocol in design
01

Current state

The OSS vector implementation stores path/content-hash metadata. Production semantic embeddings and retrieval are planned work.

output
# Current: metadata-only index
neb vector index --snapshot-id <id> --remote origin
# Records: path, content_hash, chunk boundaries
# Does NOT generate embeddings
02

Index structure

Vector indexes are keyed by TreeSnapshotId or ProjectionId, not by Git SHA. This enables version-aware retrieval.

output
VectorIndexManifest
  registry_repo_id
  snapshot_id
  index_version
  embedding_model     # Future: model identifier
  chunk_count

CodeEmbeddingChunk
  manifest_id
  path
  line_range
  content_hash:algorithm:digest
  embedding          # Future: actual vectors
03

Search

Search across indexed snapshots using the registry API or CLI.

output
neb vector search "authentication middleware" --manifest-id <id> --remote origin
04

Policy-aware retrieval

The policy layer determines which chunks are visible to each actor and environment. Vector search results respect policy decisions.