Ingestion and Storage Modes¶

Overview¶

After the LLM extracts entities and relations from your documents, Flexible GraphRAG can store the resulting knowledge graph in:

• Property Graph (Neo4j, FalkorDB, ArcadeDB, Memgraph, …) — LPG model, Cypher queries
• RDF Graph Store (Fuseki, GraphDB, Oxigraph) — triple store, SPARQL queries
• Both simultaneously — set both PG_GRAPH_DB and RDF_GRAPH_DB to non-none values

Configuration¶

Use the two picker vars to control where extracted knowledge graph data is written:

PG_GRAPH_DB=neo4j        # neo4j | arcadedb | falkordb | ... | none
RDF_GRAPH_DB=fuseki      # fuseki | graphdb | oxigraph | none

# Base namespace for entity instance URIs in RDF output
RDF_BASE_NAMESPACE=https://integratedsemantics.org/flexible-graphrag/kg/

Storage Modes¶

Property Graph Only (Default)¶

Set RDF_GRAPH_DB=none. Stores extracted entities and relations in the configured property graph store only.

PG_GRAPH_DB=neo4j
RDF_GRAPH_DB=none

Data flow:

Documents → LLM Extraction → PropertyGraphIndex → Neo4j (Cypher)

RDF Graph Store Only¶

Set PG_GRAPH_DB=none. Extracted entities and relations are converted to RDF and pushed directly to the selected RDF store. No property graph is written.

PG_GRAPH_DB=none
RDF_GRAPH_DB=fuseki
FUSEKI_BASE_URL=http://localhost:3030
FUSEKI_DATASET=flexible-graphrag

Data flow:

Documents → LLM Extraction → KGToRDFConverter → Fuseki / GraphDB / Oxigraph (SPARQL)

Use case: Semantic web / linked data workflows where you only need SPARQL access.

Both Simultaneously¶

Set both PG_GRAPH_DB and RDF_GRAPH_DB to non-none values. Extraction runs once; the same nodes are written to both destinations.

PG_GRAPH_DB=neo4j
RDF_GRAPH_DB=fuseki
FUSEKI_BASE_URL=http://localhost:3030
FUSEKI_DATASET=flexible-graphrag

Data flow:

                              ┌→ PropertyGraphIndex → Neo4j (Cypher)
Documents → LLM Extraction ──┤
                              └→ KGToRDFConverter  → Fuseki / GraphDB / Oxigraph (SPARQL)

Use case: Best of both worlds — fast Cypher traversal + SPARQL reasoning in a triple store.

Skipping Graph Extraction¶

Both the property graph and RDF graph stores can be bypassed at ingest time, while leaving previously extracted data intact for search and Q&A.

Per-document: skip_graph¶

Pass skip_graph=true on a per-ingest call to skip KG extraction (and therefore skip both PG and RDF writes) for that document only. Available from:

When skip_graph=true: - LLM KG extraction is skipped entirely (faster ingest) - No new triples are written to the RDF graph store - No new nodes/edges are written to the property graph store - Vector and full-text (BM25/search) stores are still updated normally

Global: ENABLE_KNOWLEDGE_GRAPH=false¶

Set in .env to disable KG extraction for all ingestion globally:

ENABLE_KNOWLEDGE_GRAPH=false

Equivalent to passing skip_graph=true on every ingest call — KG extraction and graph writes are skipped for both PG and RDF stores.

Previously extracted data is preserved¶

In both cases, previously ingested graph data is not deleted. Hybrid Search and AI Query / Chat will still return results from graph extractions that happened before skip_graph was used or ENABLE_KNOWLEDGE_GRAPH was disabled. To remove stale graph data, use:

• scripts/rdf_cleanup.py clear-doc <ref_doc_id> — remove one document's RDF triples
• scripts/rdf_cleanup.py clear-all — wipe the entire RDF named graph
• scripts/cleanup.py — clear the property graph store

RDF Representation¶

Entity → RDF triples¶

# rdf:type from entity label
:alice_johnson  rdf:type   onto:EMPLOYEE .

# rdfs:label from entity name (original casing preserved)
:alice_johnson  rdfs:label "Alice Johnson" .

# Datatype properties
:alice_johnson  onto:age   "30"^^xsd:integer .
:alice_johnson  onto:email "alice@example.com" .

Relation → RDF 1.2 Annotation¶

LPG relations are typed edges that carry properties:

(Alice)-[:WORKS_FOR {since: 2020, role: "Engineer"}]->(TechCorp)

This maps to RDF 1.2 inline annotation syntax (Turtle 1.2, W3C Recommendation), which is supported natively by all three stores:

# RDF 1.2 annotation syntax (default: RDF_ANNOTATION_SYNTAX=rdf_1.2)
:alice_johnson  onto:works_for  :techcorp
    {| onto:since  "2020"^^xsd:string ;
       onto:role   "Engineer"^^xsd:string |} .

The {| |} block desugars to an anonymous reifier node linked via rdf:reifies:

[] rdf:reifies <<( :alice_johnson onto:works_for :techcorp )>> ;
   onto:since "2020"^^xsd:string ;
   onto:role  "Engineer"^^xsd:string .

Annotation syntax options (set via RDF_ANNOTATION_SYNTAX):

Store support: | Store | rdf_1.2 | rdf_star | flat | |---|---|---|---| | Apache Fuseki (Jena 5) | ✅ | ✅ | ✅ | | Ontotext GraphDB 10+ | ✅ | ✅ | ✅ | | Oxigraph 0.4+ | ✅ | ✅ | ✅ | | Any SPARQL 1.1 store | ❌ | ❌ | ✅ |

URI Generation¶

Entity and predicate URIs are generated from names using a slug strategy:

When USE_ONTOLOGY=true, the ontology IRI is used as the ONTO_NS so extracted types align with the loaded OWL ontology.

RDF Store Connection Configuration¶

Configure connection details for the store selected by RDF_GRAPH_DB:

# Apache Fuseki
RDF_GRAPH_DB=fuseki
FUSEKI_BASE_URL=http://localhost:3030
FUSEKI_DATASET=flexible-graphrag
FUSEKI_USERNAME=admin
FUSEKI_PASSWORD=admin

# Ontotext GraphDB
RDF_GRAPH_DB=graphdb
GRAPHDB_BASE_URL=http://localhost:7200
GRAPHDB_REPOSITORY=flexible-graphrag
GRAPHDB_USERNAME=admin
GRAPHDB_PASSWORD=admin

# Oxigraph (HTTP mode)
RDF_GRAPH_DB=oxigraph
OXIGRAPH_URL=http://localhost:7878

See env-sample.txt (RDF / Graph Store Configuration section) for the full reference.

Terminology¶