Skip to content

Environment Configuration Guide

This document explains how to configure Flexible GraphRAG using environment variables and configuration files.

Configuration Files

File Purpose
.env Your main configuration file — copy from flexible-graphrag/env-sample.txt
flexible-graphrag/env-sample.txt Template with all options and inline examples

env-sample.txt Structure

The configuration file opens with DB selection and framework config, followed by per-service settings:

Section 1: DB Selection

PG_GRAPH_DB=neo4j               # Property graph: neo4j | arcadedb | falkordb | ladybug | memgraph |
                                 #   nebula | neptune | neptune_analytics | arangodb | apache_age |
                                 #   cosmos_gremlin | hugegraph | surrealdb | tigergraph | spanner | none
RDF_GRAPH_DB=none               # RDF store: fuseki | oxigraph | graphdb | neptune_rdf | none
VECTOR_DB=qdrant                # Vector: qdrant | elasticsearch | opensearch | chroma | milvus |
                                 #   weaviate | pinecone | postgres | lancedb | neo4j | none
SEARCH_DB=elasticsearch         # Search: bm25 | elasticsearch | opensearch | none

Section 2: Framework Config

CHUNKER_BACKEND=llamaindex      # llamaindex | langchain
GRAPH_BACKEND=llamaindex        # llamaindex | langchain (auto for LC-only stores)
VECTOR_BACKEND=llamaindex       # llamaindex | langchain
SEARCH_BACKEND=llamaindex       # llamaindex | langchain
KG_EXTRACTOR_BACKEND=llamaindex # llamaindex | langchain
RETRIEVAL_FUSION=llamaindex     # llamaindex | langchain (EnsembleRetriever/RRF)

NLTK data for the LlamaIndex chunker (NLTK_DATA)

#NLTK_DATA=./nltk_data          # optional: where SentenceSplitter finds NLTK's sentence tokenizer and stopwords

LlamaIndex's SentenceSplitter (CHUNKER_BACKEND=llamaindex, the default) uses NLTK data that ships inside the llama-index-core package. When that package was installed with uv (which hardlinks installed files to its cache), nltk 3.10.4 and later refuse to open it: Security Violation [pathsec.open]: refusing multiply-linked file ... nltk_cache\corpora\stopwords\english.

So when NLTK_DATA is not set, the backend (and the Langflow components) make a plain copy of that data once, in the user cache folder (flexible-graphrag/nltk_data/<llama-index-core version>), and use it. Nothing to configure. The backend, Langflow flow mode and the CocoIndex pipeline (also cocoindex update from the command line) all do this.

Set NLTK_DATA only to use a folder of your own (a relative path is from the folder the backend starts in): an empty folder is fine, the data is downloaded into it on first use (needs internet access once).

Sections 3+: Service-specific settings (LLM, embeddings, graph DBs, vector DBs, etc.)

Database Configuration Patterns

Use {TYPE}_*_DB_CONFIG — takes precedence over generic fallbacks:

# Property graph
PG_GRAPH_DB=neo4j
NEO4J_GRAPH_DB_CONFIG={"uri": "bolt://localhost:7687", "username": "neo4j", "password": "password"}

# Vector
VECTOR_DB=qdrant
QDRANT_VECTOR_DB_CONFIG={"host": "localhost", "port": 6333}

# Search
SEARCH_DB=elasticsearch
ELASTICSEARCH_SEARCH_DB_CONFIG={"hosts": ["http://localhost:9200"], "index_name": "hybrid_search"}

# RDF
RDF_GRAPH_DB=fuseki
FUSEKI_BASE_URL=http://localhost:3030
FUSEKI_DATASET=flexible-graphrag

Individual Variables (Legacy)

NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=password
ELASTICSEARCH_URL=http://localhost:9200
QDRANT_HOST=localhost
QDRANT_PORT=6333

Easy Database Switching

# Current: OpenAI + Qdrant + Elasticsearch + Neo4j
LLM_PROVIDER=openai
VECTOR_DB=qdrant
SEARCH_DB=elasticsearch
PG_GRAPH_DB=neo4j

# Switch to: Ollama + Milvus (LC backend) + OpenSearch + ArangoDB (LC-only)
LLM_PROVIDER=ollama
VECTOR_DB=milvus
VECTOR_BACKEND=langchain
SEARCH_DB=opensearch
PG_GRAPH_DB=arangodb
# GRAPH_BACKEND auto-set to langchain for arangodb

Configuration Best Practices

Development: Start with the defaults — Neo4j for graph, Qdrant for vector, Elasticsearch for search.

Vector-only RAG (no KG extraction, faster ingest): 

PG_GRAPH_DB=none
ENABLE_KNOWLEDGE_GRAPH=false
VECTOR_DB=qdrant

Ingest and UI defaults (v0.9.0+): three optional settings, true when not set. env-sample.txt (graph database section) sets DEFAULT_ENABLED_GRAPH=false and DEFAULT_ENABLED_RDF=false, so new rows start with search + vector only and a row's graphs are added when wanted; a REST or MCP ingest that doesn't pass enabled_graph / enabled_rdf then builds no graphs either. The integration test harnesses set both to true for their own backends.

Setting What it decides
DEFAULT_ENABLED_GRAPH Whether an ingest that doesn't say (enabled_graph not given) builds the property graph (PG_GRAPH_DB); where the Processing tab's Graph column starts for new rows
DEFAULT_ENABLED_RDF The same for the RDF graph (RDF_GRAPH_DB, enabled_rdf) and the RDF column
DEFAULT_RECURSIVE Where the Deep checkbox of the Alfresco, Nuxeo and Box forms starts

The UIs read them from GET /api/ui-defaults (see the auto-sync API reference).