What the engine is for.

The target deployment is the ERP.ai stack: thousands of apps and hundreds of thousands of agents doing graph traversal, vector retrieval, provenance, and durable write-back through one engine. Everything below is ordered by how directly it serves that.

Production use cases

ERP.ai, Neo Agent, Proto data path. The shared data path for thousands of apps and hundreds of thousands of agents: app state, agent memory, tenant context, graph relationships, vector retrieval, provenance, query history, and durable write-back in one engine. Engine path: ffsd query/load commands + Cypher + HNSW + qlog/WAL + snapshots and export/import.

Agent fleet query layer. Runtime reads for large agent fleets that constantly ask: what does this app know, which entities are connected, what changed, which source produced it, and what should be retrieved next? Engine path: indexed Cypher reads + relationship expansion + vector candidate lookup + metrics/profile paths.

Operational write-back systems. ERP.ai-style workflows where approvals, disputes, procurement, finance, customer records, tool calls, and model outputs continuously read, derive, index, and write back state. Engine path: MVCC read + durable mutation batch + secondary/vector index maintenance + qlog receipt + reopenable commit.

High-throughput ingest. Load and replay app events, agent traces, entity updates, tool outputs, node files, edge files, and mutation logs without an application sidecar owning correctness. Engine path: whole-file preflight + endpoint checks + manifest loads + checkpointed resume + qlog status.

Graph + vector retrieval with permissions. Retrieve the right context under tenant, app, role, workflow, and graph predicates. The eval harness shows why post-filtering vectors can be fast but wrong under sparse predicates — 11% recall at 1% selectivity. Engine path: predicate-aware ANN + graph expansion + typed filters + top-k recall checks.

Feature, model, and audit workflows. Keep feature views, sampled neighborhoods, embeddings, model-derived facts, source links, confidence scores, and audit trails next to the operational data they came from. Engine path: DenseF32 feature views + sampled CSR neighborhoods + model-derived writes + source/model provenance edges.

The retrieval-correctness numbers behind permissioned retrieval →