epistemic-graph 2.21.0

Rust-based epistemic graph engine for agent-utilities

epistemic-graph

One durable, Rust-native engine that is a multi-modal analytical database — graph · SQL · vector · RDF/OWL · time-series · key-value/blob — behind one query planner and one durable store.
_{Every modality is a first-class view over one RowSet algebra, from a Raspberry Pi to a replicated Raft cluster, from one core.}

Full documentation: the architecture, per-interface guides, deployment recipes, the tier/binary map, and the concept registry live at the official docs site. Start with the capability & parity matrix — every capability is tracked operation-by-operation (✅ supported · 🔶 in-progress · 🗺 roadmap). If a doc and the code disagree, the code wins.

---

What it is

A modern data platform usually needs a graph database and a vector index and a SQL warehouse and an RDF triple-store + reasoner and a time-series DB and a blob store and a full-text index — plus a broker, an observability stack, a GIS engine, and an LLM KV-cache. That is a dozen systems, a dozen copies of the data, brittle sync pipelines, and an application layer that stitches results back together.

epistemic-graph collapses that rack into one durable engine with one unified query planner. Every modality is a view over the same RowSet algebra, so a single plan can seed candidates from an OWL inference or a SPARQL pattern, filter them with SQL, traverse the graph, re-rank by vector similarity and BM25 text, fuse the results, join a time-series and a lakehouse table, and run a sandboxed WASM or numeric UDF — without ever leaving the engine or marshalling rows back to the client.

It also speaks the wire protocols of the systems it replaces — Postgres, MySQL, MSSQL, SQLite, Neo4j Bolt, Redis, S3, AMQP/MQTT/STOMP, PromQL/OTLP — so existing clients, drivers, BI tools and ORMs connect unmodified, all resolving to ONE exec path over ONE store.

It runs independently

epistemic-graph is a self-contained database. It has no dependency on agent-utilities or any agent framework — any client, in any language, over any of its wire protocols, can use it directly:

# Point DBeaver / psql / a JDBC app at the Postgres wire and just use SQL:
psql -h 127.0.0.1 -p 5433 -U agent -d epistemic

CREATE TABLE metrics (id TEXT PRIMARY KEY, value DOUBLE PRECISION);
INSERT INTO metrics VALUES ('cpu', 0.42);
SELECT id, value FROM metrics WHERE value > 0.1;

It is greatly enhanced by agent-utilities

epistemic-graph is also the compute & storage engine for agent-utilities, which exercises every modality — using the graph + OWL/RDF layer as its ontology-driven knowledge graph, the vector/text surfaces for hybrid retrieval, the agent-memory primitives for durable memory, the broker/streams for dispatch, and the KV-cache tier under vLLM/LMCache. If you run the full ecosystem, agent-utilities turns this engine into a reasoning substrate; if you don't, it is still a complete, durable, multi-modal database on its own. See the agent-utilities Graph Engine guide.

---

Why it matters

• One store, one transaction, one security model. A graph mutation + a vector upsert + a blob reference land in one redb WriteTransaction — all modalities commit together or none do. One snapshot, one ACID boundary, one per-agent RLS model, one planner. → Master-of-all engine
• Durable by default. Built redb-authoritative: the persist directory is the source of truth and an acked write survives kill -9 (commit-before-ack). → Service mode
• Scales by configuration, not by rewrite. The same binary runs embedded in-process, as a single durable server, or — with the opt-in cluster layer — as a multi-node Raft cluster with cross-shard transactions. → One build, opt-in layers · Cluster deployment
• Drop-in wire compatibility. Existing Postgres/Neo4j/Redis/S3/AMQP/PromQL clients connect unmodified. → Connecting (per-wire guide)
• One full-featured build. cargo build is the whole engine — every main feature that compiles without a GPU/robotics toolchain, in one published wheel; cluster (HA raft) and full-extras (GPU/ROS2) are opt-in build layers on top. Runs on Raspberry Pi 4+. → One build, opt-in layers
• Measured to win the agent-memory workload. Against a conventional stitched stack (separate vector DB + BM25 + app-level fusion, no KV cache, no warm-fork), the unified engine matches recall (1.000) while retrieving ~3.6× faster, reusing cross-modal context across a fan-out with retrieval_calls == 1 (vs N), keeping writes read-fresh in 25.7 ms (incremental, not full-rebuild), and surviving a full restart with a durable KV cold-tier (100% survival, >300× vs recompute). → Benchmarks

---

Quick start

Docker

docker volume create eg-data
docker run -d --name epistemic-graph \
  -e GRAPH_SERVICE_AUTH_SECRET="$(openssl rand -hex 32)" \
  -e EPISTEMIC_GRAPH_PGWIRE_ADDR=0.0.0.0:5433 \
  -p 127.0.0.1:9100:9100 -p 127.0.0.1:9101:9101 -p 127.0.0.1:5433:5433 \
  -v eg-data:/var/lib/epistemic-graph/data \
  <registry>/epistemic-graph:<tag>

The server refuses to start without GRAPH_SERVICE_AUTH_SECRET (HMAC-SHA256 on the RPC transport). --allow-insecure opts out for development only. Full recipes (compose, HA cluster, prebuilt wheels): deployment guide.

Binary

# start a complete single-node database with the Postgres wire listener
EPISTEMIC_GRAPH_PGWIRE_ADDR=127.0.0.1:5433 \
  epistemic-graph-server --persist-dir /var/lib/eg

Python client

pip install epistemic-graph

from epistemic_graph import SyncEpistemicGraphClient

g = SyncEpistemicGraphClient()                    # connects/attaches to the UDS engine
g.nodes.add("AgentA", {"type": "coordinator"})
g.edges.add("AgentA", "AgentB", {"weight": 1.5})
print("Order:", g.graph.topological_sort())

More entry points — the remote → shared-local → autostart resolver and the embedded in-process handle — are in engine modes.

---

Capabilities

Each surface has a deep-dive; the capability & parity matrix is the operation-by-operation source of truth.

Query surfaces & interfaces

Surface	You can point at it…	Deep dive
SQL / pgwire	psql, DBeaver, JDBC/ODBC, BI tools, ORMs — DataFusion SELECT (joins/CTE/window), full DML, arbitrary user tables + DDL + COPY + ALTER TABLE, CREATE FUNCTION incl. PL/pgSQL bodies, pg_catalog/information_schema, pgvector/AGE/Timescale/ParadeDB compat. pgwire is in the one main build.	SQL & pgwire
SPARQL / RDF / OWL	Stardog/GraphDB clients — SPARQL 1.1 SELECT/ASK/CONSTRUCT/DESCRIBE/UPDATE, the W3C /sparql endpoint, OWL 2 EL⁺/RL and DL-tableau + SWRL, SHACL/ShEx + ICV write-path enforcement, R2RML, GeoSPARQL	SPARQL & RDF · Ontology lifecycle
Cypher / Bolt	Neo4j drivers & cypher-shell — MATCH/writes/WITH/aggregation, GDS via CALL gds.*, native Bolt v4.4 wire	Cypher & Bolt
GraphQL	Apollo clients — Federation v2 subgraph, subscriptions over CDC, fragments/variables/directives, relay pagination, APQ/depth/cost hardening	GraphQL
Vector / ANN	Pinecone/Milvus-style kNN — IVF-PQ + OPQ + SQ8 + HNSW + exact/flat, hybrid metadata pre-filter, cross-shard scatter-gather, real pgvector ANN pushdown	Vector / ANN
Time-series	InfluxDB/Timescale-style — time_bucket, ASOF, gap-fill, OHLC, decay, SQL window frames, Op::Window planner op	Time-series
Key-value & Blob	Redis (RESP2/3 + pub/sub + MULTI/EXEC) and S3/MinIO (REST, multipart, range GET); content-addressed streaming CAS with content-defined chunking; embedded in-process handle	Key-value & Blob
Messaging & Broker	RabbitMQ/Kafka clients — exchanges/routing, DLQ, TTL, priority, delayed delivery, consumer groups, publisher confirms + exactly-once; AMQP/MQTT/STOMP wires; replayable streams	Messaging & Broker
Observability	Prometheus/OpenObserve/Jaeger clients — log ingest, PromQL (extended fn set), OTLP traces, VRL pipelines, federated search; the engine also emits its own OTLP + Prometheus remote-write	Observability
GIS / Spatial	PostGIS-style — CRS/reprojection, R-tree, GeoJSON/WKB/GPX + Shapefile/KML/GeoParquet, MVT + raster tile pyramids, routing (turn-restrictions/time-windows)/isochrones/TSP	GIS / Spatial
KV-cache (LLM)	vLLM/LMCache — tiered hot/warm/cold KV-block cache (zstd/lz4) + shared dedup backend + HTTP endpoint: the durable L2 tier under vLLM's GPU cache and LMCache's CPU tier	KV-cache
Agent memory	Zep/mem0/LeanRAG-style — bi-temporal AsOf, summary tier, episodic→semantic consolidation, decay/reinforce, hierarchical retrieval, scene/trajectory — all drivable over the wire	Agent memory
Clients	Python (full) · JS / Go (thin) over framed MessagePack, no PyO3/FFI	Client drivers

The unified UQL planner is what lets these compose in a single cross-modal plan; natural-language → query (NlQuery) is a complete, LLM-optional seam.

Analytics & advanced capabilities

• Analytics Program — "one kernel, two surfaces." A BLAS/LAPACK-free Rust numeric kernel (eg-numeric, faer + ndarray) exposed as (A) an in-process Python extension + numpy-shim and (B) in-database DataFusion UDFs/UDAFs — cosine_sim/l2_normalize/zscore/covariance scalars + svd/pca/kmeans column→matrix aggregates, and the differentiator: cross-modal join → analytics in-engine (join graph ⋈ vector ⋈ time-series, then run pca/kmeans over the joined set — impossible in numpy, which has no data layer).
• Lakehouse LTAP interop. The engine's tables materialize as open Parquet + Delta + Iceberg (real Iceberg v2 Avro manifests, per-column stats for predicate pushdown) with an Iceberg-REST catalog + LSN as-of, so Databricks/Spark/Trino/DuckDB read them with zero ETL; every materialize/compact/delete run also emits a real OpenLineage RunEvent (optional HTTP push), feature lake.
• Distribution / Robotics / GPU tail. Cross-region async read-replicas, Calvin deterministic commit, ROS2 bridge (rosbridge-WS + pure-Rust RTPS), and a GPU distance/tensor dispatch seam with a real CUDA backend + CPU fallback.

Distribution & durability

• redb-authoritative, commit-before-ack (kill -9-safe), folded into every tier.
• In-engine Raft replication (cluster tier, openraft) with automatic failover; off ⇒ byte-for-byte single-node.
• Cross-shard 2PC (presumed-abort + parallel-commit + read-only-participant + non-blocking Raft-replicated decision), multi-Raft groups + online resharding, and an opt-in Calvin deterministic-ordering branch.
• One placement authority. An epoch'd PlacementCatalog (cluster/raft) drives online split/merge/move via a prepare-then-fenced-cutover sequence — a caller holding a stale epoch is redirected, never served stale — and takes priority over the hash-ring router for any graph with an explicit placement entry. Paired with real multi-group production startup (EPISTEMIC_GRAPH_RAFT_GROUPS) and cross-shard read fan-out.
• Durable analytics-job plane (eg-jobs, opt-in feature jobs) — Method::AnalyticsJob async submit/status/cancel/resume over an immutable input-snapshot handle (a graph + pinned OCC version, re-readable later via AS OF), with results committed through the same claim/evidence convention the epistemic layer reads.
• Lazy graph lifecycle (EPISTEMIC_GRAPH_LAZY_STARTUP/EPISTEMIC_GRAPH_MAX_RESIDENT_GRAPHS) — catalog-only boot + a bounded resident-graph cache, both off/unbounded by default.
• Cross-modal ACID across all modalities in one transaction.

→ Engine scaling program · Multi-Raft status · Catalog-driven resharding

Epistemic reasoning (opt-in — eg-epistemic)

The RowSet result surface can carry engine-native Claim/Evidence/BeliefState with cycle-guarded confidence propagation and a BELIEF AS OF/EVIDENCE FOR/CONTRADICTS UQL family (feature epistemic); a paraconsistent truth-maintenance engine (Dung grounded/preferred/stable argumentation) that now auto-invalidates dependents on a committed write via a real server-side CDC hook, plus the bitemporal epistemic_status why/why-not/what-changed capstone (feature epistemic-tms); and policy-aware proof redaction — ExplainBelief's disclosure_level masks (never silently drops) an evidence node the caller's row-level-security context can't see (feature epistemic-redaction). A multimodal evidence graph (EvidenceSpan, 11 located-evidence locus kinds spanning text/table/image/audio/video/metric/SQL-row/code/trace) is reachable under epistemic; its blob-CAS-backed citation resolver lives behind the separate opt-in alignment feature. Calibrated causal reasoning (a linear-Gaussian structural causal model with genuine Pearl do-calculus — observe/intervene/counterfactual) is implemented and tested inside eg-epistemic, but is not yet wired onto the server/wire protocol — an engine-native library capability today, not a servable operation. None of this is on by default; every feature above is additive on top of the full build. See docs/capabilities.md.

Security & isolation

• Mandatory auth — every RPC carries an auth token. The default remains v0 (HMAC-SHA256(secret, request_id)); an opt-in v1 signed envelope (EG-P0-5) additionally binds method/graph/tenant/principal/body-hash/timestamp/nonce under one HMAC, constant-time verified — off by default (EPISTEMIC_GRAPH_REQUIRE_SIGNED=1 to require it; v0 keeps working unchanged until then). Transport TLS/mTLS and OIDC principals are not part of this yet. pgwire separately adds SCRAM-SHA-256. See Service mode for the full contract.
• Per-agent Row-Level Security applied before any query surface touches the graph; the result cache keys on the caller's RLS context. The default posture remains permissive/back-compat — an unowned/undecodable/untagged-legacy row stays visible to all, unchanged from every pre-EG-P0-6 deployment. An opt-in strict/default-deny posture (EG-P0-6, EPISTEMIC_GRAPH_RLS_DEFAULT_DENY=1) denies such a row unless it's explicitly _visibility: "public" or _owner-tagged — set it to get the stricter posture; it is not the shipped default.
• Encryption-at-rest (ChaCha20-Poly1305, pure-Rust) + a hash-chained tamper-evident audit log covering every one of the ~80 gateway-routed mutating methods (the single MutationPlan/ commit_mutation gateway) plus a documented, machine-checked-empty triage of every other mutating method (registry-lifecycle/cluster-admin/process-global/txn-self-routed — EG-P0-2).

→ Service mode

Interface	Operation	Status	Feature	Notes
SQL	SELECT (joins, aggregates, CTE, window, subquery)	✅	query	DataFusion 43 over nodes + edges; real predicate pushdown (Inexact)
SQL	INSERT / UPDATE / DELETE (+ RETURNING)	✅	query	nodes + user tables (EG-KG.query.follow-up); serializable CAS gates
SQL	Compound/AND/OR/IN/BETWEEN/IS NULL WHERE DML, INSERT…SELECT, UPDATE…FROM/DELETE…USING, ON CONFLICT upsert	✅	query	EG-045/046/047/048; serializable re-check under the write guard
SQL	Mixed-store wire transactions (BEGIN/COMMIT/ROLLBACK + TransactionStatus)	✅	pgwire	EG-KG.compute.kg-transaction-is-pinned; node + user-table ops, read-your-own-writes; documented non-2PC user-table window
SQL	CREATE VIEW/DROP VIEW, CREATE FUNCTION, arrays/ranges + common functions	✅	query	EG-072/118/104; durable view + function catalog
SQL	Arbitrary user tables + DDL (CREATE/ALTER ADD COLUMN/DROP), COPY	✅	query	durable redb table catalog (EG-KG.query.register-user-tables-alongside/EG-KG.query.register-each-user-table); JOINable to the graph
SQL	ALTER TABLE beyond ADD COLUMN — DROP/RENAME COLUMN, RENAME TO, ALTER COLUMN TYPE, DROP CONSTRAINT	✅	query	durable catalog rewrite with data migration (EG-KG.query.rename-table-moves-catalog)
SQL	Columnar segments + window functions (ROW_NUMBER/RANK/LAG/LEAD/OVER(…))	✅	query	EG-089; struct-of-arrays analytical scan
Postgres compat	pg_catalog + information_schema system views (\d/\dt/\l)	✅	pgwire	EG-KG.query.route-create-view-create; synthesized from live table/view/function catalogs
Postgres compat	CREATE EXTENSION catalog · pgvector vector + <->/<=>/<#> + real ANN pushdown to HNSW/IVF + exact re-rank	✅	pgwire	EG-KG.query.create-drop-extension-over/115/116; real top-k pushdown (EG-KG.query.real-pgvector-ann-top)
Postgres compat	AGE cypher() set-returning function, TimescaleDB hypertables + continuous aggregates, ParadeDB @@@ real BM25 ranking + snippets	✅	pgwire	EG-KG.query.postgres-family-extension-plan/117/119; real BM25 (EG-311)
Postgres wire	listener, simple + extended/prepared protocol	✅	pgwire	EPISTEMIC_GRAPH_PGWIRE_ADDR; in the one main build (EG-KG.compute.capability-reference/EG-KG.sharding.deployment-tiers)
Postgres wire	SCRAM-SHA-256 / trust auth, pg_catalog introspection	✅	pgwire	EG-KG.query.concept-13 / EG-KG.query.datafusion; pg user → engine ACL actor
SPARQL	SELECT (BGP, paths, FILTER subset, OPTIONAL, UNION, GROUP/agg, BIND, DISTINCT, SLICE)	✅	sparql	spargebra parser compiled to LPG scans
SPARQL	ASK / CONSTRUCT / DESCRIBE	✅	sparql	template instantiation + bounded description (gated by rdf, implied by sparql)
SPARQL	UPDATE (INSERT/DELETE DATA, DELETE/INSERT WHERE, CLEAR, CREATE/DROP GRAPH)	✅	sparql	eg-rdf/src/update.rs; LOAD intentionally deferred (no HTTP fetch in write path)
SPARQL	/sparql HTTP endpoint (W3C SPARQL 1.1 Protocol)	✅	sparql-http	src/server/sparql_http.rs; GET + POST query/update
SPARQL	true named graphs (quad dataset) + FROM/FROM NAMED	✅	sparql	GRAPH ?g/constant-IRI over registry graphs (EG-KG.ontology.from-from-named)
SPARQL	ORDER BY total-ordering, VALUES, MINUS, EXISTS/NOT EXISTS, negated property set	✅	sparql	EG-KG.ontology.completing-eg-order-by/125/055/056; fixes the unordered-results correctness gap
SPARQL	content negotiation (JSON/XML/CSV/TSV/Turtle/N-Triples), rich FILTER, sub-SELECT, SERVICE federation	✅	sparql	EG-KG.ontology.content-negotiation-serializers/053/051/052; SSRF allowlist on SERVICE
SPARQL	SHACL + ShEx validation, ICV integrity constraints (enforced on the commit/write path), GeoSPARQL + RCC8/Egenhofer	✅	sparql/geosparql	EG-KG.ontology.concept-6/133/146/261/155; ICV commit-guard (EG-KG.ontology.rdf-update-guard)
RDF I/O	JSON-LD 1.1, TriG, N-Quads, RDF/XML serialization matrix	✅	rdf	EG-KG.ontology.eg-concrete-syntax-matrix/137 (alongside Turtle/N-Triples)
Cypher	MATCH … WHERE … RETURN … LIMIT (var-length [*m..n])	✅	cypher	read over a snapshot; WHERE supports OR/IN/STARTS WITH/CONTAINS/IS NULL (EG-KG.query.eg-extend-read-side)
Cypher	writes (CREATE/MERGE/SET/DELETE+DETACH/REMOVE)	✅	cypher	native eg-core mutations (EG-KG.query.cypher-execution)
Cypher	ORDER BY/SKIP/WITH/OPTIONAL MATCH/OR/aggregation/DISTINCT/UNWIND/CALL	✅	cypher	EG-KG.query.eg-extend-read-side/141/142; GDS via CALL gds.<algo>(…) YIELD … streams eg-compute results as rows (EG-KG.query.eg-2/144/298)
Cypher	Neo4j Bolt v4.4 wire (PackStream v2)	✅	bolt-wire	EPISTEMIC_GRAPH_BOLT_ADDR (EG-KG.query.bolt-wire-protocol); neo4j drivers / cypher-shell
GraphQL	read queries (scan + BFS, schema-from-graph, aliases, first/limit, filters)	✅	graphql	byte-equal to Cypher path
GraphQL	mutations (createNode/updateNode/deleteNode/addEdge/removeEdge)	✅	graphql	native eg-core mutations
GraphQL	Apollo Federation v2 subgraph (_service/_entities, @key) + APQ/depth/complexity hardening	✅	graphql	EG-295/296
GraphQL	subscriptions / fragments / variables / directives / relay pagination	✅	graphql	real CDC push (LiveQuery, WS/SSE) + $/@ lexer + fragments + @skip/@include + relay envelope (EG-064/065/066)
OWL	EL⁺ + RL forward-chaining materialization & classification	✅	owl	pure-Rust; consistency + incremental + justifications
OWL	confidence-weighting + Ebbinghaus time-decay	✅	owl	EG-KG.ontology.concept-13; per-axiom eg:confidence, fact decay
OWL	query-time Op::Reason (reasoner seeds a RowSet)	✅	owl-plan	distributed/cross-shard union supported
OWL	OWL-DL (tableau, cardinality, allValuesFrom), SWRL user rules	✅	owl-dl/owl	pure-Rust DL tableau (consistency→classification→instance) + swrlb: built-in library; EL/RL fast path stays default (EG-KG.ontology.concept-2/060)
Vector / ANN	IVF-PQ + OPQ + SQ8-refine, persistent (reopen w/o rebuild), warm-on-start	✅	ann	parallel/SIMD brute-force fallback below threshold
Vector / ANN	hybrid metadata pre-filter (kNN + allow(id) predicate)	✅	ann	search_filtered (EG-070); filtered during the ADC probe
Vector / ANN	HNSW index (higher recall-per-probe than IVF-PQ)	✅	ann	insert/search/serde-persist, recall-harness-tuned (EG-KG.retrieval.hnsw-vector-index)
Vector / ANN	exact/flat kNN index + ANN-vs-exact re-rank + recall@k/precision harness	✅	ann	EG-KG.query.concept-5
Vector / ANN	cross-shard kNN scatter-gather → deterministic global top-k	✅	ann	server-layer scatter over per-shard indexes merged via the merge_topk leaf (EG-319, completing EG-KG.retrieval.scatter-gather)
Time-series	store + time_bucket, ASOF join, gap-fill LOCF, OHLC, downsample, decay	✅	tsdb	native redb columnar, no DataFusion
Time-series	time-ops as unified planner ops (Op::Window)	✅	tsdb	real window_aggregate over the RowSet via eg-tsdb time_bucket (EG-KG.query.streaming-execution); per-point retention trim (EG-KG.temporal.bucket-cutoff-trim)
Blob / CAS	content-addressed streaming store (redb-native)	✅	blob	refcount mark-and-sweep GC; bounded RAM
Blob / CAS	S3 / MinIO backend behind the same ChunkStore trait	✅	blob-s3	manifest/linkage byte-identical
Blob / CAS	content-defined chunking	✅	blob	Gear/FastCDC rolling-hash chunker (variable boundaries); sha256 CAS dedup + refcount GC preserved (EG-KG.storage.backward-manifest-read)
Key-value	embedded in-process engine API over redb rows	✅	embedded	EmbeddedEngine — no Tokio/socket/HMAC (EG-KG.backend.engine-modes)
Key-value	generic namespaced get/put/scan/cas KV surface over redb	✅	redb	src/server/kv.rs (EG-022); durable, commit-before-ack; not graph-scoped
Multi-wire	wire-neutral SQL core (WireProtocol/WireSession, one classify→exec path)	✅	wire	src/server/wire (EG-KG.compute.subsystems-reference); shared by every SQL wire
Multi-wire	MySQL / MariaDB wire (handshake v10 + mysql_native_password)	✅	mysql-wire	EPISTEMIC_GRAPH_MYSQL_ADDR (EG-KG.query.kg-2)
Multi-wire	MSSQL TDS wire	✅	mssql-wire	EPISTEMIC_GRAPH_MSSQL_ADDR (EG-KG.query.hand-rolled-tds-server)
Multi-wire	SQLite-dialect NDJSON-over-TCP endpoint	✅	sqlite-wire	EPISTEMIC_GRAPH_SQLITE_ADDR (EG-KG.query.concept-3); .db file I/O 🔶 forward roadmap
Multi-wire	Neo4j Bolt v4.4 wire (PackStream v2, native Cypher)	✅	bolt-wire	EPISTEMIC_GRAPH_BOLT_ADDR (EG-KG.query.bolt-wire-protocol)
Broker	exchanges (direct/topic/fanout) + bindings/routing over the EG-KG.compute.atomically-claim-oldest-pending work-queue	✅	broker	RabbitMQ-class (EG-275)
Broker	DLQ · message/queue TTL · priority · delayed/scheduled delivery · consumer-groups + QoS/prefetch	✅	broker	EG-KG.compute.dead-letter-queues/277/278/279/280
Broker	replayable append-log streams (Kafka-style offsets/retention) + publisher confirms + manual ack/nack	✅	broker	EG-283/284
Broker	exactly-once (idempotent-producer dedup) + stream/confirm/ack over AMQP confirm.select + MQTT 5 frames	✅	broker	effectively-exactly-once atop at-least-once (EG-314)
Broker wires	AMQP 0.9.1 · MQTT 3.1.1/5.0 · STOMP 1.2 listeners	✅	amqp-wire/mqtt-wire/stomp-wire	EPISTEMIC_GRAPH_{AMQP,MQTT,STOMP}_ADDR (EG-275/281/282)
KV / structures	Redis RESP2/3 wire (strings/hashes/lists/sets/sorted-sets) + pub/sub + MULTI/EXEC	✅	redis-wire	EPISTEMIC_GRAPH_REDIS_ADDR (EG-KG.ontology.resp2-resp3-codec-round/307)
Object store	S3-compatible REST (bucket/object CRUD, SigV4-lite) over the blob CAS + multipart upload + range GET	✅	s3-api	EG-KG.ontology.object-put-get-head/307
Observability	log ingest + PromQL /api/v1/query (extended fn set) + OTLP traces /v1/traces + service-dependency map	✅	obs/promql/traces	EPISTEMIC_GRAPH_OBS_ADDR, default :5080 (AU-KG.ingest.self-ingest/172/163); PromQL _over_time/topk/label_replace/clamp* (EG-302)
Observability	VRL-style ingest pipelines (parse/filter/enrich, cross-modal) + super-cluster federated search (typed SQL/SPARQL fusion)	✅	obs/federation	EG-165/243; typed result fusion (EG-KG.query.schema-typed-fusion-sql)
Observability	engine's own telemetry egress — OTLP metrics/traces export + Prometheus remote-write receiver	✅	otel-export	the engine emits, not just ingests (EG-316)
Lakehouse (LTAP)	Parquet-on-object-store + Delta log + Iceberg-REST catalog + LSN as-of — external lakehouse engines read with zero ETL	✅	lake	eg-lake (EG-KG.storage.lsn-as-snapshot-returns); Iceberg Avro manifest is a documented stub
Spatial / GIS	SpatialScan + ST_Within/ST_DWithin, GeoSPARQL + RCC8/Egenhofer, CRS/reproject, R-tree, GeoJSON/WKB/GPX + Shapefile/KML/GeoParquet	✅	geo/geosparql	eg-geo (EG-KG.ontology.singles-concept/261/155/262/263/264); Shapefile/KML/GeoParquet I/O (EG-KG.domains.geo-formats); no GEOS/PROJ
Spatial / GIS	XYZ/TMS + Mapbox Vector Tiles · weighted routing (+ turn-restrictions/time-windows)/isochrones/TSP · map-based task tracking	✅	geo	EG-KG.domains.map-tiles/266/267; turn-restrictions + time-dependent weights (EG-KG.domains.geo-partitioning)
Document / JSON	deep JSONPath query + durable inverted path-index (persists to redb + boot-rehydrate + cost Stats); PG ->/->>/@> + Mongo $match	✅	(core)/query	Pred::JsonPath (EG-084); durable path-index (EG-308)
Tensor / probabilistic	N-D array store (CAS-backed) + TensorScan/TensorOp (results written back to the CAS store); distribution-valued properties	✅	tensor	EG-085/086; tensor-op CAS write-back (EG-304)
Scene-graph / 3D	:SceneObject pose + transform hierarchy + spatial relations (robotics/AR/urban-3D)	✅	(core)	EG-087
CEP / streams	windowed event ingest + Op::Cep bounded-NFA pattern match over sliding/tumbling windows	✅	stream	EG-KG.query.pipelined-execution
Robotics	multimodal sensor fusion (ASOF-aligned) + action/trajectory memory	✅	tensor	EG-KG.query.multi-rate-sensor-stream/099
KV-cache (LLM)	tiered hot/warm/cold KV-block cache (real zstd/lz4 warm-tier compression) + shared dedup backend + HTTP endpoint (vLLM/LMCache contract) — the durable L2 tier under vLLM (L0 GPU) → LMCache (L1 CPU)	✅	kvcache-server	eg-kvcache (EG-185/186/187); real compression codec (EG-315); see kvcache interface
Agent memory	bi-temporal AsOf, decay/reinforce, summary-node tier, episodic→semantic consolidation, LeanRAG retrieval, scene/trajectory — drivable over the wire	✅	(core)	Op::AsOf (AU-KG.compute.kg-2); EG-KG.compute.hierarchical-summary-tier-eg/221/222/195; wire-Op surface (EG-KG.memory.eg-batch-decay-caller)
Epistemic substrate	belief/confidence propagation + bitemporal why/why-not/what-changed + policy-aware proof redaction; opt-in: multimodal-evidence citation resolver (Method::ExplainEvidence) + calibrated causal do-calculus intervention + provenance-aware retrieval ranking (Method::CausalEstimate/RankByProvenance)	✅	epistemic(+epistemic-tms/epistemic-redaction/evidence-graph/epistemic-causal)	CONCEPT:EG-KG.epistemic.epistemic-substrate; CONCEPT:EG-X1 (multimodal citation); EPI-P3-3 (causal + ranking) — all facade-reachable, none folded into full
OBDA	R2RML virtual graphs — SPARQL over a foreign source rewrites to ForeignScan (no materialization); parses R2RML Turtle documents	✅	federation	EG-101; R2RML Turtle parse (EG-305)
RBAC	durable roles + role hierarchy + resource/action grants over per-agent RLS (persist to redb + boot-reload)	✅	security	EG-KG.compute.feature; durable persistence (EG-KG.compute.durable-rbac-identity-persistence)
Scheduling / QoS	real-time QoS/SLO scheduler — per-tenant/priority admission + deadline scheduling + backpressure	✅	(server)	EG-320; complements the reserved read-admission lane (EG-KG.coordination.reserved-read-lane)
Backup / DR	consistent online backup + restore CLI + PITR (Method::Backup/Restore)	✅	redb	EG-090
Full-text	Tantivy BM25 inverted index, RankText + reciprocal-rank fusion	✅	text	composes in the unified planner
Unified planner	Scan·Filter·Traverse·Rank·RankText·FuseRrf·Reason·SparqlBgp·Udf·ForeignScan·AsOf·Limit	✅	query+	each op feature-gated; see UQL
Unified planner	Op::Window / Op::Foreign execution	✅	query	Op::Window = real eg-tsdb windowed aggregate (timeseries, EG-KG.query.streaming-execution); Op::Foreign resolves the name via the ForeignSourceRegistry (federation, EG-073)
UQL	text DSL → wire::Plan (one parse, zero new exec path)	✅	(front-end always ships)	dependency-free parser
UQL	natural-language → query (Method::NlQuery, /nl, nl_query() UDF)	✅	nl-query	EG-078/080; complete LLM-optional seam — inert until an OpenAI-compatible endpoint is set; AU-provider integration tracked on the agent-utilities side
Durability	redb-authoritative, commit-before-ack (kill -9-safe)	✅	redb	folded into every tier
Distribution	openraft replication + automatic failover	✅	raft	cluster tier; off ⇒ byte-for-byte single-node
Distribution	cross-shard 2PC + parallel-commit + read-only-participant + non-blocking (Raft-replicated decision) commit	✅	raft	presumed-abort 2PC (EG-KG.storage.lane-n-increment) + parallel prepare/empty-write-set skip (EG-KG.txn.cross-shard) + Paxos-Commit-lite replicated decision (EG-KG.txn.harness-crash)
Distribution	Calvin deterministic-ordering cross-shard commit (order-first, vote-free, crash-replay)	✅ opt-in	calvin (⇒ nonblocking)	EG-KG.txn.calvin-deterministic-ordering; sequencer + Raft-replicated input log + vote-free execution shipped; distributed OLLP read-lock phase + multi-node epoch fan-in documented-deferred — see distribution-robotics-gpu
Distribution	multi-Raft groups (N-group ring, online reshard, hibernate/rehydrate)	✅	raft	GroupRouter + MultiRaft (KG-2.266/267/268); online ownership move
Distribution	cross-region async read-replica tier (bounded-LSN log + follower pull → wal::apply) + capacity guardrails (circuit breaker / per-tenant quota / backpressure)	✅ opt-in	federation-search	EG-322/323; off by default (EPISTEMIC_GRAPH_REPLICATE); complements the EG-320 QoS scheduler with absolute ceilings
Federation	remote engine / HTTP-JSON / external SQL (sqlx) as a ForeignScan	✅	federation(-sql)	in the main build; activates only when a foreign source is registered
Robotics	ROS2 bridge over rosbridge-WebSocket (CDC↔topic, no DDS/C toolchain)	✅ opt-in	ros2-bridge	EG-KG.domains.robotics-gpu-distribution; pure-Rust tokio-tungstenite
Robotics	native DDS/RTPS ROS2 wire — TWO legs behind one DdsTransport trait, both zero-config rmw-mangled	✅ opt-in	ros2-dds (pure-Rust rustdds) / ros2-rmw (real CycloneDDS-C, S5)	EG-KG.ingest.dds-transport / EG-KG.ingest.rmw-cyclonedds-leg; ros2-rmw vendors + cmake-builds the CycloneDDS C sources (no network/libclang at build time) for genuine live-ros2 interop
GPU	GPU distance/tensor dispatch seam + real CUDA backend (NVRTC, dynamic-loading)	✅ opt-in	gpu / gpu-cuda	EG-KG.compute.gpu-distance-seam/327; pure-Rust CPU backend is always compiled + the byte-for-byte ground truth; CUDA auto-falls-back on a GPU-less host; live-GPU kernel validation deferred (none in CI)
Numeric / analytics	BLAS/LAPACK-free Rust numeric kernel (eg-numeric: reductions/stats · element-wise · linalg via faer · seedable random)	✅ P1 (Surface A) · 🗺 Surface B	pip install epistemic-graph[numeric] (Surface A) · cargo numeric/full (Surface B)	AU-KG.compute.numeric-kernel; P1 of the Analytics Program ("one kernel, two surfaces"). Surface A = in-process Python — the kernel .so is folded into the one epistemic-graph wheel as epistemic_graph.numeric (no separate eg-numeric package; + the AU xp numpy-shim, 847 parity checks); Surface B (in-DB DataFusion UDFs/graph/vector/ts analytics) is P2–P5 follow-on. Surface-B numeric cargo feature is part of the main build — see numeric-kernel
Clients	multi-language client drivers — Python (full) · JS / Go (thin: broker/streams/RBAC/backup/NL) over framed MessagePack (no PyO3/FFI)	✅	(client)	EG-328; wire-parity gated (test_protocol_parity.py) — see clients

---

Architecture at a glance

flowchart TB
    subgraph Clients["Clients (unmodified)"]
        AU["agent-utilities / graph-os"]
        PG["psql / DBeaver / BI / ORM (pgwire)"]
        DRV["Neo4j · Redis · S3 · AMQP · PromQL drivers"]
        EMB["Embedded in-process caller (Pi/edge)"]
    end

    subgraph Engine["epistemic-graph-server (one Rust process)"]
        WIRE["Wire adapters (one WireProtocol exec path)"]
        SEC["Security: per-agent RLS + audit chain + encryption-at-rest"]
        PLAN["Unified RowSet planner (cost-reordered, cross-modal)"]
        CORE["GraphCore: petgraph + ledger + result cache"]
        MOD["Modalities (feature-gated): vector · SQL · RDF/OWL · TSDB · text · blob · GIS · tensor · stream"]
        DUR["Durability: redb-authoritative + Raft + cross-shard 2PC + CDC"]
    end

    AU --> WIRE
    PG --> WIRE
    DRV --> WIRE
    EMB --> CORE
    WIRE --> SEC --> PLAN --> CORE
    CORE --> MOD
    CORE --> DUR

See technical overview for the crate DAG and the planner pipeline, and subsystems for how each modality composes on the one store.

---

One build, opt-in layers

The engine ships as one prebuilt binary — the full-featured build — plus two opt-in build layers. The target host pulls a prebuilt wheel and never compiles. Full map: one build, opt-in layers.

Build	Carries	For
main (default wheel)	the whole single-node DB: redb-authoritative + cypher + DataFusion SQL + graphql + ann + tsdb + blob + Tantivy text + rdf/sparql/owl + wasm-udf + security + federation + the wire family (pgwire/mysql/mssql/sqlite/bolt/redis/amqp/mqtt/stomp) + the numeric kernel + kvcache-server + broker	anything single-node, Raspberry Pi 4+ to workstation
+ cluster	main + Raft replication + cross-shard 2PC + distributed compute	multi-node HA
+ full-extras	main + accelerator legs (gpu-cuda, ros2-bridge, ros2-dds)	GPU / robotics

cluster and full-extras are opt-in build flags layered on top of the one main build (openraft is cluster-only; cudarc/rustdds are full-extras-only) — they are built from source, not published as separate wheels.

---

Documentation

• Capabilities & parity matrix — the operation-by-operation truth table.
• Technical overview · Master-of-all engine · UQL & the unified planner.
• Interfaces: Connecting · SQL · SPARQL · Cypher · GraphQL · Vector · Time-series · KV & Blob · Messaging · Observability · GIS · KV-cache · Agent memory · Ontology · Clients.
• Analytics & distribution: Analytics Program · Analytics in UQL · Numeric kernel · Lakehouse LTAP · Distribution / Robotics / GPU.
• Deploy & operate: Standalone deployment · DBeaver / psql quickstart · Deployment topology · Tiers & binaries · Deployment · Engine modes · Service mode · Cost model & capacity · Operations runbook.
• Contributing: AGENTS.md · Rust compute guide · CONTRIBUTING.md.

Design for a network boundary. Every out-of-process call is a serialize → socket → deserialize round trip, not a function call. Batch, never per-element: ship work as one op over data already in the graph, and keep tight per-element math in-process. See the Rust compute guide.

License

MIT — see LICENSE.