AssetOps Knowledge Graph

12,647 nodes. 12,662 edges. IBM AssetOpsBench at 99% accuracy -- deterministic graph queries, zero LLM tokens.

Part of the Samyama ecosystem — loaded into and queried via the graph engine at samyama-ai/samyama-graph. This repo holds the loader and source-data specifics for the KG.

Demo

A narrated walkthrough (load the ISO 14224 + ISA-95 asset graph → most-critical equipment → DEPENDS_ON failure-impact propagation → triage of priority-1 work orders raised by high-severity anomalies):

python -m demo.demo                                                            # run live
asciinema rec --overwrite --cols 92 --rows 32 --idle-time-limit 2.0 \
  -c "bash -c 'source ~/projects/venv/bin/activate && PYTHONUNBUFFERED=1 python -m demo.demo'" \
  demo/assetops.cast                                                           # re-record
agg demo/assetops.cast demo/assetops.gif                                       # convert

Generation-Augmented Knowledge (GAK)

When the graph doesn't have an asset type — here an electric motor, absent from the chiller+AHU graph — the engine's LLM agent writes the missing failure modes into the graph as provenance-tagged nodes (source:"LLM-derived"), which the re-query then answers deterministically. It's the inverse of RAG (write structured facts in, vs. retrieve text out) — Architecture D from the VLDB 2026 paper.

python -m demo.demo_gak                                                         # run live (needs SGE + an enrich LLM on the tenant)
agg demo/assetops_gak.cast demo/assetops_gak.gif                                # convert

---

IBM's GPT-4 agents score 65% on their own AssetOpsBench using flat document stores. We loaded the same data into a knowledge graph and asked:

"What equipment is affected if Chiller 6 fails?"

MATCH (e:Equipment {name: 'Chiller-6'})<-[:DEPENDS_ON*1..3]-(downstream:Equipment)
RETURN downstream.name, downstream.criticality_score
ORDER BY downstream.criticality_score DESC

Equipment	Criticality
AHU-3	0.92
CRAC-2	0.88
AHU-7	0.85

The DEPENDS_ON topology and criticality_score shown above are an analytical layer we add on top of IBM's data (paper §3.2), so this query runs on the extended graph; the base graph loaded directly from IBM's sources has 9 node labels and 5 edge types. Across the 139 IBM scenarios, an instrumented run shows 86 deterministic answers come from a live graph query and 53 from domain-knowledge handlers — see docs/information-leakage-analysis.md.

137/139 scenarios passing. 63ms average. Zero tokens. The bottleneck was the data model, not the LLM. Powered by Samyama Graph.

---

Results

Approach	Pass Rate	Avg Latency	Tokens
GPT-4 + flat docs (IBM)	91/139 (65%)	not reported	not reported
GPT-4 + graph NLQ	114/139 (82%)	~5,800 ms	~4,600/scenario
Deterministic (graph)	137/139 (99%)	63 ms	0

Same model (GPT-4), same data, +17pp improvement -- proving the gain comes from the data model.

Schema

9 node labels -- Equipment, Sensor, FailureMode, WorkOrder, Location, Site, Event, AnomalyEvent, AlertEvent

5 edge types -- CONTAINS_LOCATION, CONTAINS_EQUIPMENT, HAS_SENSOR, FOR_EQUIPMENT, MONITORS

Data source -- IBM AssetOpsBench (139 scenarios, 9 asset classes)

Quick Start

Load from snapshot (recommended)

# Download (475 KB)
curl -LO https://github.com/samyama-ai/samyama-graph/releases/download/kg-snapshots-v5/assetops.sgsnap

# Start Samyama and import
./target/release/samyama
curl -X POST http://localhost:8080/api/tenants \
  -H 'Content-Type: application/json' \
  -d '{"id":"assetops","name":"AssetOps KG"}'
curl -X POST http://localhost:8080/api/tenants/assetops/snapshot/import \
  -F "file=@assetops.sgsnap"

Build from source and benchmark

git clone https://github.com/samyama-ai/assetops-kg.git && cd assetops-kg
git clone https://github.com/IBM/AssetOpsBench.git ../AssetOpsBench
pip install -e ".[dev]"
python -m benchmark.run_ibm_scenarios --data-dir ../AssetOpsBench   # 99%
python -m benchmark.run_samyama                                      # 100%

Example Queries

-- Dependency chain: what breaks if this equipment fails?
MATCH (e:Equipment {name: 'Chiller-6'})<-[:DEPENDS_ON*1..3]-(downstream:Equipment)
RETURN downstream.name, downstream.criticality_score
ORDER BY downstream.criticality_score DESC

-- Failure modes monitored by sensors
MATCH (s:Sensor)<-[:HAS_SENSOR]-(e:Equipment)<-[:MONITORS]-(fm:FailureMode)
RETURN e.name, fm.name, s.type, fm.severity
ORDER BY fm.severity DESC

Links

Samyama Graph	github.com/samyama-ai/samyama-graph
The Book	samyama-ai.github.io/samyama-graph-book
IBM AssetOpsBench	github.com/IBM/AssetOpsBench
Contact	samyama.dev/contact

License

Apache 2.0