GraphRAG Configuration Guide - Complete Property Reference

backendUrl

String

"http://conversation-service:8080"

The internal network endpoint for the primary Conversation microservice.

backendSseMessageUrl

String

"${backendUrl}/sse/message"

Formatted internal destination path used to dispatch streaming Server-Sent Events (SSE) messages back to the active client.

internalN8nUrl

String

"http://n8n:5678"

The internal network address used by services to coordinate background loopbacks within the orchestration engine.

keycloakUrl

String

"https://identity.example.com/auth"

The external, browser-facing target path where clients authenticate OAuth2/JWT tokens.

keycloakRealm

String

"chatbot-app"

Target security realm partition designated for your application users and configurations.

primaryLLMModel

String

"openai.gpt-4.1"

The primary LLM engine is utilized across non-MCP processing phases (e.g., intent categorization, summary extraction, and final generation).

secondaryLLMModel

String

"openai.gpt-4.1-mini"

The fallback model engine activates automatically if the primary hosting region goes offline.

mcpLLMModel

String

"anthropic.claude-3-5-sonnet-20240620-v1:0"

A specialized model optimized for analyzing graph structures, generating SPARQL queries, and parsing triples.

vectorSearchUrl

String

"${graphRagComponentsUrl}/vector/search"

Internal cluster network endpoint pointing directly to your deployed Components Service vector search module.

vectorSearchTopK

Number

5

Specifies the exact number of nearest-neighbor text chunk results ($k$-NN) returned from the vector index per retrieval cycle.

vectorStorePreset

String

"opensearch_native"

Selects the active target vector provider preset logic. Supported options:

vectorStoreCustomMappins

Object

{}

Key-value schema overrides applied when vectorStorePreset is explicitly configured to "custom".

vectorIndex

String

""

The targeted index or collection ID. Required when using isolated standalone engines like Pinecone, Qdrant, or Weaviate.

vectorStoreMetadataPassthrough

Boolean

true

When set to true, it forwards raw source metadata fields to response payloads for downstream citation or verification.

vectorStoreMetadataExclude

Array

["content"]

Identifies specific verbose or sensitive properties to strip from the forwarded metadata object to optimize payload transit.

embeddingsProvider

String

"aws"

Identifies the hosting environment provider for vector generation (e.g., "aws", "openai", or "custom").

embeddingsModelId

String

"cohere.embed-english-v3"

Defines the specific model architecture signature used to transform clean text fragments into dense vectors.

graphModelingServerUrl

String

"https://graph-modeling.example.com"

Base access endpoint address of your running taxonomies and modeling platform.

graphModelingProjectId

String

"00000000-0000-0000-0000-000000000000"

The unique validation project UUID assigned to your enterprise taxonomy partition.

conceptProperties

Array

Standard SKOS array

A collection of SKOS properties or custom URIs fetched dynamically to extract and contextualize domain definitions.

graphDBMcpUrl

String

"http://graphdb:7200/mcp/sse"

The target network access endpoint for the GraphDB Model Context Protocol (MCP) server.

graphDBMcpRepository

String

""

The main repository namespace inside your database instance holding semantic triples.

graphDBMcpOntologyGraph

String

""

Named graph URI enclosing the ontologies, schemas, and relational maps of your dataset.

graphDBMcpFtsSearchTriplesLimit

Number

100

Caps the maximum volume of triples extracted through full-text search indexing to protect the orchestrator's token limits.

shortMemoryMaxUncompressedSizeInTokens

Number

300

The token volume limit threshold. If an active session exceeds this limit, automatic context compression and summarization are triggered.

parallelStepsMaxTiemout

Number

65

Caps parallel execution phases inside the orchestrator in seconds.

datasetContext

String

""

Optional

A short, plain-language description of your knowledge base content passed to the output guardrails to evaluate generation in light of specialized domain boundaries.