Home / Solutions / Evidence Graph
PRODUCT · TRACEABLE QUESTION ANSWERING · KG-RAG

Evidence Graph

Answers about your documents, with every statement connected to the exact source text that supports it. Short, direct answers, with the evidence shown below.

99%
of extracted facts correct (held-out docs)
87%
of expected facts found (held-out docs)
Fielded with NATO ACT
SHINE 2026, ACT Innovation Continuum, Halifax. On-site experiment, integrated via MCP.

Where usual RAG falls short

  • Unsupported statements. The answer can say things no passage supports.
  • No line of evidence. It cannot show which passage supports which statement.
  • Scattered facts. Facts spread across many documents are hard to collect.

How it works

01 · INGEST

Fixed versions

Documents split into passages with exact positions.

02 · EXTRACT

Against your ontology

Accepted only with an exact supporting quote.

03 · RESOLVE

Merges cite evidence

Uncertain items stay out until reviewed.

04 · RELEASE

Versioned

Graph and index released together and recorded.

05 · RETRIEVE

Hybrid search

Keywords, meaning and graph links.

06 · CHECK

Sentence by sentence

Unsupported sentences removed before the answer is shown.

The evidence chain

Every statement goes back to its source: answer sentence → cited passage → graph claim → exact quote → document version and position → release.

Tested on a corpus built to trip up an AI

On the Arctic Fracture NATO wargame scenario, 51 synthetic reports covering the opening 30 hours of an Article 5 response deliberately test provenance, identity and uncertainty. Evidence Graph's conflict and uncertainty checks now answer these correctly, each with its evidence chain:

Trick question
Answer
Why it is hard
Does the corpus prove the provisional contact ALDER is hostile?
No
Suspicion is not proof
Were five carrier aircraft transferred to task group 3 command?
Not established
Support is not a transfer of command
Does warning notice WARN-01 show an actual launch?
No
A warning is not an event

Arctic Fracture is a fictional exercise scenario; all countries, forces and events in it are fictional. The test reports are synthetic.

Architecture

Access
Web interface · REST API · MCP server for AI agents
Question service
Plans the search; searches text, vectors and graph; reranks; writes and checks the answer
Report processing
Splits reports into passages, extracts records and links, checks matches, keeps sources
Ontology & review
Allowed record, link and attribute types; uncertain matches held for review
Storage
PostgreSQL + pgvector for passages, vectors and audit history; Neo4j for records, claims and source links
Models
Locally hosted language and vector models, or cloud providers; a separate model for extraction and match checks

Honest scope

Single facts, conflicts and unanswerable questions are reliable today. Complete lists and counts, time-dependent and multi-step questions are being improved.