epistemic-graph 0.28.0

Rust-based epistemic graph engine for agent-utilities

epistemic-graph

Unified Rust-Native Compute Engine for AI Agent Infrastructure
_{Consolidates graph operations, quantitative finance, data science, AST analysis, and OWL reasoning into a single high-performance binary.}

Documentation — The architecture, service-mode operations, Rust compute reference, measured transport benchmarks, and concept registry for the engine are maintained in the official documentation.

This is the compute engine for agent-utilities — a standalone Rust service reached out-of-process over MessagePack/UDS (no PyO3). You can use it on its own (binary + pure-Python client), or let agent-utilities drive it. Contributing? See CONTRIBUTING.md.

---

Architecture

The epistemic-graph crate is the singular computation engine for the agent-utilities ecosystem. All high-performance operations route through this crate, exposed to Python out-of-process over a long-running Tokio service speaking length-prefixed MessagePack over Unix Domain Sockets (default) or TCP, authenticated with HMAC-SHA256. There is no PyO3 / in-process FFI — the engine runs as a separate process (maturin ships it as bindings = "bin"), so callers cross a network boundary, not a function call. This is enforced by scripts/check_no_pyo3.sh.

┌──────────────────────────────────────────────────────────┐
│                    epistemic-graph                        │
│                                                          │
│  ┌─────────┐  ┌──────────┐  ┌─────────┐  ┌───────────┐  │
│  │  Graph   │  │ Finance  │  │  Data   │  │ Reasoning │  │
│  │  Core    │  │ Engine   │  │ Science │  │  Engine   │  │
│  │(petgraph)│  │          │  │         │  │ (Datalog) │  │
│  └─────────┘  └──────────┘  └─────────┘  └───────────┘  │
│  ┌─────────┐  ┌──────────┐  ┌─────────┐                 │
│  │   AST   │  │ Semantic │  │  Algo   │                 │
│  │ Parser  │  │  Store   │  │ Library │                 │
│  └─────────┘  └──────────┘  └─────────┘                 │
│                                                          │
│  ╔════════════════════════════════════════════════════╗   │
│  ║  Tokio Server (UDS/TCP + HMAC-SHA256 auth)        ║   │
│  ╚════════════════════════════════════════════════════╝   │
│           ↕ length-prefixed MessagePack (UDS / TCP)      │
│  ╔════════════════════════════════════════════════════╗   │
│  ║  Python: EpistemicGraph class                     ║   │
│  ╚════════════════════════════════════════════════════╝   │
└──────────────────────────────────────────────────────────┘

Features

Core Graph Engine (CONCEPT:KG-2.2)

• petgraph-backed: Native-compiled Rust graph structures
• Temporal Knowledge Graph (TKG): Ebbinghaus Forgetting Curve and fact decay natively integrated
• Topological Sort: Sub-millisecond DAG resolving
• DFS Cycle Detection: Returns precise cycle paths
• Shortest Path: Efficient unweighted BFS traversal
• Blast Radius: Transitive impact analysis to configurable depth
• PageRank & PPR: Centrality computation
• Community Detection: Louvain-style graph clustering
• VF2 Subgraph Isomorphism: Pattern matching queries
• Reactive State Ledger: Transaction log with replay for backend persistence

Finance Engine (CONCEPT:QF-1.0)

• Portfolio Optimization: Mean-variance (MVO), min-variance, risk-parity, efficient frontier
• Risk Metrics: VaR (historical + Monte Carlo), CVaR, Sortino, Calmar, max drawdown
• Regime Detection: Hidden Markov Model (Baum-Welch + Viterbi)
• Signal Generation: Rolling Z-score, EWMA, momentum, alpha combination, information coefficient
• Execution Algorithms: TWAP/VWAP scheduling, market impact estimation, LOB matching
• Pairs Trading: Spread signal generation and regime-aware position sizing

Data Science Engine (CONCEPT:DS-1.0)

• OLS Regression: Gradient descent with configurable learning rate and epochs
• K-Means Clustering: Parallel centroid computation
• PCA: Eigenvalue decomposition via power iteration
• Dataset Statistics: Mean, std, min, max, correlation matrix
• Estimators: ridge / lasso / elasticnet / decisiontree / randomforest / gradientboosting / adaboost / svr (replaces sklearn on the hot path)
• Training loss / optimizer kernels (CONCEPT:KG-2.22): softmax / log_softmax, cross_entropy (+grad), dpo_loss (Bradley-Terry, +grads), grpo_surrogate (PPO clip, +grad), kl_divergence (Schulman k3), adam_step / sgd_step — the pure-Rust performance path for the in-house training substrate, mirroring data-science-mcp trainers/objectives.py. client.datascience.{...}.

Reasoning Engine (CONCEPT:KG-2.23)

• Transitive/Symmetric Inference: Compiled Datalog closures
• Domain/Range Rules: OWL-style type inference
• Property Chain Composition: Multi-hop rule chaining

AST Parser

• Multi-Language: 9 languages via tree-sitter — Python, Rust, TypeScript, JavaScript, Go, Java, C, C++, C# (src/parser/tree_sitter.rs::lang_for_path)
• Full Granularity: Functions, classes, methods, imports stored as Symbol nodes
• Repository Ingestion: Directory walker with automatic graph population

---

Cargo Feature Flags

Feature	Description	Dependencies
ast	Tree-sitter AST parser	tree-sitter, language grammars
finance	Quantitative finance engine	(pure Rust)
datascience	ML primitives	(pure Rust)
reasoning	OWL/Datalog reasoning	(pure Rust)
compute	All compute features	finance + datascience + reasoning
server	Tokio UDS/TCP server	tokio, hmac, rmp-serde
full	Everything	compute + server + ast
all	Alias for full	(same as full)

Build examples

# Library only (default, no server)
cargo build

# With compute modules
cargo build --features compute

# Full build including server binary
cargo build --features full

# Run all tests
cargo test --lib --features compute

---

Quickstart

1. Installation

uv pip install -e .
# or
pip install -e .

2. Python Usage

The client speaks to the out-of-process engine and exposes capabilities through typed namespaces (g.nodes, g.edges, g.graph, g.finance, g.datascience, g.reasoning, ...). Use SyncEpistemicGraphClient for blocking code or EpistemicGraphClient for async.

from epistemic_graph import SyncEpistemicGraphClient

# Connect to the running engine (starts/attaches to the UDS service)
g = SyncEpistemicGraphClient()

# Graph operations
g.nodes.add("AgentA", {"type": "coordinator"})
g.nodes.add("AgentB", {"type": "worker"})
g.edges.add("AgentA", "AgentB", {"weight": 1.5})

print("Order:", g.graph.topological_sort())
print("Cycle:", g.graph.find_cycle())

# Finance — portfolio optimization
weights = g.finance.optimize_portfolio([0.1, 0.15, 0.08], [[0.04, 0.01, 0.005], ...], 0.02)

# Finance — risk metrics
metrics = g.finance.risk_metrics([0.01, -0.02, 0.03, -0.005, 0.02])

# Data Science — regression
coeffs = g.datascience.linear_regression([[1.0, 2.0], [3.0, 4.0]], [3.0, 7.0])

# Reasoning — OWL/RDFS forward chaining (CONCEPT:KG-2.17)
# Materialises inferred edges/types in-graph and returns the inferred triples.
result = g.reasoning.reason(
    subclass_relations=[("Dog", "Animal")],
    transitive_properties=["ancestor"],
)
print("Inferred:", result["inferred_count"], "triples")

---

Security: Authentication & Tenant Isolation

• Auth is mandatory. Every request carries an HMAC-SHA256(secret, request_id) token. The server refuses to start with an empty secret — set GRAPH_SERVICE_AUTH_SECRET (or --auth-secret). To intentionally run unauthenticated (development only) pass --allow-insecure or set EPISTEMIC_GRAPH_ALLOW_INSECURE=1; the server then starts with a prominent SECURITY: warning naming the bind addresses.
• ACLs are enforced in dispatch. Once any identity is registered (client.consensus.register_identity), every graph-targeted operation is checked by the isolation layer (src/isolation.rs): peer agent graphs are denied, managers reach subordinate graphs, team graphs are member-read / manager-write, global: graphs are read-only, and __commons__ stays open to all authenticated agents. Callers identify themselves with the optional agent_id field (EpistemicGraphClient.connect(..., agent_id="worker1")). With zero registered identities nothing is checked — single-tenant deployments are unchanged. Violations return ACCESS_DENIED: ... errors.
• TCP has no TLS. The optional TCP listener is plaintext; keep it on loopback or behind a TLS-terminating proxy / WireGuard / SSH tunnel.

See docs/service_mode.md for the full protocol, policy table, and examples.

---

Engine Internals

Capabilities living in the binary beyond the headline features:

• HNSW semantic store (src/compute/semantic.rs) — per-graph embedding store with an hnsw_rs approximate-nearest-neighbor index for O(log n) cosine search, falling back to brute force below 32 embeddings. Served over the protocol as AddEmbedding / SemanticSearch; search results are re-weighted by temporal confidence decay (Ebbinghaus) before ranking.
• Lock-free heavy compute (CONCEPT:KG-2.51) — CPU-heavy read-only operations (semantic search, PageRank, betweenness, community detection, MST, similarity edges, VF2, lifecycle metrics) never run while holding the per-graph lock: dispatch takes a cheap structural snapshot (GraphCore::topology_snapshot / analysis_snapshot) under the read lock and computes on the tokio blocking pool, so a large analytics request no longer stalls writers on that graph. Single-pass O(V+E) ops stay under-lock (a snapshot would cost as much as the computation).
• Prometheus metrics (src/metrics.rs, cargo feature metrics, on by default; CONCEPT:KG-2.51) — per-op request counters + latency histograms, in-flight / admission-permit gauges, BUSY-rejection counter, per-graph op counters and node/edge gauges (bounded label cardinality), checkpoint duration/timestamp, and auth-failure / ACL-denial counters. Exposed by a dependency-free HTTP listener on --metrics-addr / GRAPH_SERVICE_METRICS_ADDR (disabled when unset; e.g. 127.0.0.1:9101), entirely separate from the MessagePack RPC transports.
• Parser symbol metadata (src/parser/tree_sitter.rs) — beyond raw symbols, each parse extracts per-symbol metadata (name, kind, line, docstring, argument list) plus import edges, and stamps every symbol with a stable language label so the graph can answer per-language queries ("all Java symbols") and compute per-language metrics.
• Spectral clustering (src/compute/spectral.rs) — a normalized-Laplacian spectral cluster navigator. Source-only today: the module is not compiled into the crate (excluded from compute/mod.rs) and the SpectralCluster protocol method returns a deprecation error pointing at the datascience primitives (kmeans/pca).
• Hypergraph interaction encoding (src/compute/hypergraph.rs) — a seeded 2-layer MLP positional-interaction encoder with an encoder cache, used to embed (position, position) interactions into fixed-width vectors. Compiled, but its HypergraphEncodeInteraction protocol method is deprecated in favor of the datascience primitives.
• Execution orchestrator (src/execution/orchestrator.rs) — scaffold for executing compiled task graphs (topological scheduling of TaskGraphSpec). Not wired into the crate (no execution module in lib.rs); orchestration currently lives in agent-utilities, which compiles task graphs client-side.

---

Scaling & HA Reality

What the architecture does and does not give you (measured numbers in docs/benchmarks.md):

• Sharding is client-side. Shards are independent server processes, one graph universe each; the Python ShardRouter (epistemic_graph/pool.py) picks a shard per graph name with rendezvous/HRW hashing over GRAPH_SERVICE_ENDPOINTS. There is no server-side coordination, rebalancing, or cross-shard query.
• No replication, no HA. Each graph lives in exactly one process's memory. If a shard dies, its graphs are unavailable until the process restarts and reloads its last snapshot.
• RPO = checkpoint interval. Durability is periodic snapshotting (--persist-dir, default every 300 s) plus checkpoint-on-shutdown; a crash loses writes since the last checkpoint. There is no WAL.
• The 100M-agent figure is a projection, not a load test: it assumes ~52 kB resident per agent (measured on bounded 40-node subgraphs), 64 GB RAM per host, and linear shard scaling — arithmetic that yields ~78 hosts.

---

Development & Test

Run Unit Tests

# Rust tests (29 compute + graph tests)
cargo test --lib --features compute

# Python tests
uv run pytest

Format and Lint

pre-commit run --all-files

---

Documentation

• Technical Overview — Rust-side structures and graph algorithm layouts.
• Concept Registry — Registered CONCEPT bridges.
• AI Agent Handbook — Quick command sheet for coding assistants.
• Changelog — Progression of updates and releases.

---

Environment Variables

Variable	Description
GRAPH_SERVICE_AUTH_SECRET	HMAC-SHA256 secret for inter-process authentication (required — the server refuses to start without it unless the insecure opt-out is set)
EPISTEMIC_GRAPH_ALLOW_INSECURE	1/true: explicit opt-out allowing an empty auth secret (development only; logs a prominent warning)
GRAPH_SERVICE_SOCKET	Path to Unix Domain Socket for UDS communication
GRAPH_SERVICE_ENDPOINTS	Comma-separated shard endpoints consumed by the Python ShardRouter
EPISTEMIC_GRAPH_MAX_INFLIGHT	Server backpressure cap (default 1024); excess load is shed with BUSY
GRAPH_SERVICE_METRICS_ADDR	Prometheus /metrics HTTP listener address (e.g. 127.0.0.1:9101). Disabled when unset
XDG_RUNTIME_DIR	Directory for UDS socket placement

License

This project is licensed under the MIT License — see the LICENSE file for details.