Ask your documents what they actually say.
Whether your corpus is project appraisals, enterprise architecture, policy libraries, or operational procedures — EdgeCore ingests it and returns exact, citation-backed answers in seconds. The graph understands your domain semantics, not just keyword proximity. Every claim traces back to the source span, and if the answer isn't in the graph, EdgeCore says so.
as_of filter is on every query. Answers reflect a point in time, not a blended average of document versions.Why graph traversal beats document search for project work
The questions project and programme teams ask every week — which risks are rated High? which component carries the most disbursement risk? what was decided in the last review? — are structurally unanswerable by document search. They require graph traversal: typed nodes, typed edges, and an ontology that knows what a Risk or a DLI actually is. Each example below uses the same question against the same document corpus; the gap is structural, not a matter of model quality.
"Which business capabilities have no supporting application?"
cap-017 · Digital Identity Management
gap_kind: unsupported
evidence: [assn-2341, assn-2342]
Complete. Every Capability node in scope, LEFT JOIN on SUPPORTS edges. No edge = no supporting application. Returns the full list, provenance on every row.
"Based on the document, it appears that Digital Identity Management and Customer Analytics may lack dedicated applications…"
Hedged. Samples prose — misses capabilities not mentioned by name. No provenance. Cannot detect absence with confidence.
"Which capabilities are at risk because their supporting application is being retired?"
TechFlow CRM → retiring 2025-06-30
→ Customer Relationship Management
gap_kind: partially_supported
Direct SUPPORTS edge. Retirement date as a sortable typed field. Pre-joined, pre-ranked by urgency. One query across two document sources.
"The document mentions that TechFlow CRM is being sunset. CRM applications typically support customer-facing capabilities."
Inference — "typically support" is not a graph edge. Cannot join data across separate inventory and lifecycle documents. Cannot sort by retirement urgency.
"Which applications integrate with more than three other systems?"
API Gateway · integration_count: 7
high_density: true
partners: [OrderService, PaymentService,
CustomerPortal, ReportingEngine…]
WHERE integration_count > 3 applied in graph traversal. Counts both source-side and target-side edges — symmetric, complete.
"Based on the integration catalogue, the API Gateway appears to integrate with several systems."
Cannot enforce "more than three." Misses integrations where the app is the target. Only sees interfaces named in prose.
"Which applications are on end-of-life technology?"
Oracle DB 11g · eol: 2024-12-31
FinanceLedger · criticality: mission_critical
risk_kind: eol · window: 18 months
eol_date < now + 18 months. Criticality join on Application node. Unknown stacks returned as a separate risk class.
"Oracle Database is mentioned as being deprecated. Some applications may depend on it."
No criticality join. Cannot filter by 18-month window from prose dates. Cannot return "unknown stack" as a finding.
"Show me data assets with PII that have no data steward assigned"
0 rows returned
Query ran. No DataAsset node with
contains_pii = true AND steward = null.
Structurally unambiguous. 0 rows = data quality gap — the ingest hasn't captured PII flags yet. That gap is itself the finding.
"Customer email and payment records likely require stewardship oversight based on standard data governance practices…"
Fabricates a plausible answer. The uncertainty is buried in hedging language — easy to miss. No audit trail. No way to distinguish this from a real finding.
Pattern summary
| Capability | Raw LLM | Graph |
|---|---|---|
| Enumeration | Sampling from prose — misses items | Complete — every node in scope |
| Aggregation (counts, sums) | Unreliable | Exact — graph traversal |
| Multi-hop joins | Inference across documents | Trivial — graph stores edges |
| Threshold filtering | Cannot enforce | Applied in Cypher WHERE clause |
| Temporal filtering | Misses date relationships | as_of temporal model |
| Provenance | None | Assertion ID on every claim |
| Failure mode | Plausible fiction | Explicit empty result |
Twenty packs. Every project domain. One query engine.
EdgeCore ships with 20+ domain packs, all focused on the work of complex projects and programmes. Each pack defines entity types, extraction templates, Cypher query templates, and a golden eval set. A new domain is a new YAML pack — no kernel changes required. The packs below are the core set for project and programme work; enterprise architecture, org design, strategy, and process packs are also available.
From documents to graph-grounded answers in four steps.
No data project. No re-platforming. Point EdgeCore at your existing documents and the kernel does the extraction. Four steps from corpus to queryable graph.
edgecore_ingest at a document (PDF, DOCX, HTML). The kernel extracts typed (subject, predicate, object) triplets using a pack template. Every triplet cites a source span from the document — anchor ID, page, character offset.valid_from, valid_to, believed_as_of, source_doc. The temporal model means you can query "as of 2024-Q4" against any corpus — and the graph knows what was true then.edgecore_plan selects the right Cypher templates from your installed packs, runs them against the graph, and returns evidence rows — each with assertion IDs that trace back to source document spans.edgecore_ask composes a natural-language answer from the evidence rows via MCP sampling. Every claim is grounded in a graph assertion. If the evidence rows are empty, the answer says so — it does not infer from general knowledge.Five short clips. The full picture.
Each clip is two to three minutes. Watch them in order for the clearest picture, or jump to the capability you care about.
Point EdgeCore at a document. The kernel extracts typed triplets and anchors every claim to a source span.
Assertions land in a typed Neo4j graph. See how the temporal model and provenance edges are structured.
edgecore_plan selects templates, runs Cypher, and returns evidence rows. The answer is grounded — no inference beyond what the graph holds.
Run a best-practice assessment. Findings and ratings attach to graph entities and trace back to source spans.
Add a new domain pack in YAML only. No kernel changes. A new answering capability in one PR.
Native MCP server. Works with any MCP client.
EdgeCore runs as an MCP server. Your AI assistant — Claude Code, Cursor, any MCP-capable client — calls edgecore_ask, edgecore_plan, edgecore_ingest, and edgecore_scope directly in the session. No context switching. No separate query tool. Graph answers arrive inline.
EdgeCore or OrgEdge — which one?
Both EdgeCore and OrgEdge address problems with organisational knowledge — but from different angles, with independent graph schemas, and no shared data model. They are separate products; the right choice depends on your use case.
- Experience
- Interactive web application — visual digital twin explorer
- Primary input
- You model the organisation through the UI — strategy, structure, people, process
- Primary output
- Living digital twin with graph visualisation, AI assessments, transformation planning
- Best for
- Leaders managing org design, strategy alignment, compliance culture
- Pricing
- $30 / position / month
- Experience
- MCP server — graph answers arrive inline in your AI coding session
- Primary input
- You ingest your existing document corpora — EA repos, PADs, policy libraries
- Primary output
- Graph-grounded answers via MCP tools, with assertion IDs on every claim
- Best for
- Enterprise architects, analysts, and program managers querying document-heavy corpora
- Pricing
- Early access program — contact us
If your team produces large volumes of architecture documents, project appraisals, or policy libraries and needs to ask cross-document questions with full audit trails — EdgeCore is the right tool. It runs independently, with its own ontology and its own graph instance.
The graph knows what your programme actually says.
EdgeCore came out of working on complex project and programme environments — World Bank PADs with layered results frameworks, programmes with dozens of risk registers, governance libraries with interlocking practice standards. The pattern is always the same: someone needs to know what's true right now, and the answer is buried across ten documents in hedged prose. With EdgeCore, that's a graph query. The ontology understands the structure of your project; the answers reflect it.