A verifiable atomic knowledge graph
Most retrieval systems hand a model a pile of text chunks and hope. Agentyk Knowledge instead builds a graph of atomic, source-cited claims — so every answer traces back to the exact sentence it came from. Audit-ready, governed, and running against your own database.
Not retrieval-as-usual
The difference between a knowledge graph you can verify and a vector index you have to trust.
Commodity RAG
- Splits documents into opaque embedding chunks
- Answers cite a chunk, not a fact
- Provenance stops at the document
- Vocabulary is whatever the model invents
- Deletes can orphan vectors behind your back
Agentyk Knowledge
- Decomposes documents into atomic claims
- Every answer cites the exact source sentence
- Full provenance: document, position, claim
- SHACL-governed YAGO / schema.org vocabulary
- Erasure cascades from claims to vectors
How it works
From raw documents to verifiable answers in three stages.
Ingest
Upload PDFs, DOCX, TXT, or Markdown. A parser pack with OCR (for scans and images) turns every source into clean text — fully asynchronous, per-tenant fair-share, with a job you can poll.
Extract claims
An extractor decomposes the text into atomic, source-cited claims and maps their entities and relations onto the governed ontology. Each claim keeps a pointer to its exact origin.
Ground & verify
Queries are answered against the claim graph and your vector store. Every answer cites the claims it used and ships a verification code that resolves to the original sources.
What makes it verifiable
Six properties that turn retrieval into an auditable knowledge layer.
Atomic
Source-cited atomic claims
Documents are decomposed into atomic, individually-checkable claims — not opaque embedding chunks. Each claim is a single assertion you can read, audit, and trace.
Provenance
Full provenance per claim
Every claim links back to the exact source sentence, document, and position it came from. Answers are audit-ready and tamper-evident by construction.
Your data
Runs against your own database
Point it at your own graph store and your own vector database (Qdrant). Your knowledge stays in infrastructure you control — nothing is pooled with other tenants.
Governed
SHACL-governed vocabulary
A pinned YAGO / schema.org base ontology, validated with SHACL shapes, plus a per-client supplement. New vocabulary is admitted provisionally and held for review — strict opt-in.
Verifiable
Verify any answer by code
Each grounded answer carries a verification code. A public /verify/{code} endpoint resolves it back to the underlying claims and sources — independent, after the fact.
GDPR
Delete cascades to vectors
Deleting claims or collections purges the corresponding vectors from your store in the same operation — closing the orphaned-vector gap and keeping erasure requests honest.
Where it fits
Anywhere a wrong answer is expensive and a citation is non-negotiable.
Build on a knowledge layer you can audit
Ground your AI in atomic, source-cited claims — hosted in the EU, running against your own database.