Knowledge graph and vector recall for AI agents
Prime is an optional module that adds knowledge graphs, vector search, and agent recall to the AllSource event store. Combine graph traversal, HNSW embeddings, provenance, and temporal context in one memory layer.
Everything an AI agent needs to remember
Knowledge Graph
Model entities and their relationships as a directed graph. BFS traversal discovers multi-hop connections across domains — customers to orders to products to suppliers — without pre-defined joins.
Vector Embeddings
HNSW index stores high-dimensional embeddings alongside your events. Semantic similarity search finds related entities even when they share no common identifiers or schemas.
Compressed Index
Auto-generated cross-domain scaffolding links entities that co-occur across event streams. A published AllSource project evaluation measured improved cross-domain recall on its test corpus; results depend on your data and queries.
Recall API
Hybrid search that combines vector similarity, graph traversal, and temporal ordering in a single query. Your AI agents get ranked results with provenance — not just embeddings, but the events that produced them.
Core-Backed Projection Reads
A concurrent in-memory map serves Core projection reads separately from event ingestion. The published 11.9us p99 Core benchmark is not an end-to-end Prime hybrid-recall measurement.
Offline / Embedded Mode
Run Prime in-process via a Rustler NIF — no network hop, no separate service. Ship durable AI agent memory inside your Elixir application or as a standalone embedded binary.
One API call. Full context.
# Ask Prime to recall relevant context for an AI agent
curl -X POST https://api.all-source.xyz/api/v1/prime/recall \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "What do we know about customer acme-corp?",
"top_k": 5,
"strategy": "hybrid"
}'
# Example response shape; scores depend on corpus and query
{
"nodes": [
{
"entity_id": "customer:acme-corp",
"type": "customer",
"score": 0.94,
"source": "graph+vector",
"relationships": [
{ "to": "order:ord-7821", "rel": "placed", "weight": 0.88 },
{ "to": "ticket:tkt-312", "rel": "opened", "weight": 0.72 }
],
"last_event": "2026-04-15T09:32:11Z"
},
{
"entity_id": "order:ord-7821",
"type": "order",
"score": 0.88,
"source": "graph",
"payload": { "total": 24500.00, "status": "fulfilled" },
"last_event": "2026-04-14T14:07:44Z"
}
],
"strategy_used": "hybrid"
}Works with the Core event store
Hosted Prime reads and writes tenant-scoped prime.* events through Core. Local Prime embeds Core in-process. In both modes, durable events are the source for graph and vector projections.
All Prime data is a projection. If you rebuild Prime from scratch, it replays events from Core and arrives at the same state. No separate source of truth.
You can run Core without Prime. The event store stands alone for event sourcing, time-travel queries, and projections. Prime is purely additive.
Hosted Prime keeps only warm per-tenant projections in memory and rebuilds them from Core after a cold start. Local Prime keeps its event log and projections in the selected data directory.
Add durable memory to your AI agents
Prime plugs into any AllSource Core deployment. Knowledge graphs, vector search, and compressed recall — no external dependencies, no separate database.