- Graphwise Platform Documentation
- GraphRAG
- GraphRAG Configuration Guide
- Configuration of Services
Configuration of Services
30/06/2026
GraphRAG version 1.2.0
The Conversations Service manages active user sessions, orchestrates streaming outputs, and maps inbound API calls to runtime workflows. It persists session and conversation histories to an embedded DuckDB database residing on its own dedicated volume.
Configuration for this service is split between the infrastructure layer (for internal service behavior) and the runtime layer (for workflow integration):
Infrastructure Layer (Helm)
Manages low-level service properties, JVM/Spring behaviors, and resource allocations. These are defined within the Helm chart under
conversation.configuration.propertiesinsidecharts/graphrag/values.yaml.Runtime Layer (n8n Workflow)
Manages service routing and environment endpoints. The n8n configuration workflow must be updated to reference this service's network location via the
backendUrlandbackendSseMessageUrlproperties.
Review the following parameters post-deployment to ensure proper connectivity and runtime behavior between layers.
The Helm properties n8n.base.urland n8n.chat.webhook.urlmust accurately resolve to your active Workflows (n8n) service.
Default chart value:
http://graphrag-workflows:5678Verification
Confirm this matches the actual service name in your cluster by running
kubectl -n graphrag get svcIncorrect mapping will cause inbound webhook calls to fail.
The n8n Configurationworkflow node must be explicitly pointed to the Conversations Service deployment.
Configuration Target
Update the
backendUrlconstant in your n8n configuration workflowValue
Typically
http://conversation[1] -service:8080(or your internal cluster/Docker service name [1]really minor and more general observation
The sse.timeout.n8n.millisecondsproperty caps how long the service will wait for an external workflow execution to complete before severing the Server-Sent Events (SSE) stream.
Default Chart Value
300000 (5 minutes)
Recommendation
Increase this value if you expect slow, multi-hop workflow runs to prevent premature truncation.
Note
Property Formatting
The infrastructure parameters use Spring-style dotted property names set under conversation.configuration.properties. Do not format them as environment variables (for instance, do not use N8N_BASE_URL). Ensure they match the exact lowercase dotted casing defined in the chart.
Storage Volume Verification
The default Helm chart specifies the embedded DuckDB database destination path as
jdbc:duckdb:/var/lib/graphrag/conversation/duckdb.db
Tip
If legacy documentation or internal templates reference /duckdbdata/graphrag-conversation.db, verify your active volume mount configuration against your current deployment to prevent path conflicts or data persistence failures.
The Components Service is a custom micro-service that extends native vector store capabilities. It processes text-to-vector embeddings generation, orchestrates downstream search executions and runs k-NN (k-nearest neighbors) queries against your configured vector store to return relevant document chunks back to the orchestrator.
To prevent synchronization errors, parameters configured across both planes must remain structurally uniform.
Infrastructure Layer (Helm)
Configures physical vector database connectivity endpoints, underlying environment authorization mapping, and network transmission sizes.
Runtime Layer (n8n Workflow)
Configures index selections, database response-parsing schemas (presets), search threshold limits, and metadata pass-through filters.
Review and configure the following parameters to align your infrastructure deployment with your runtime orchestrator's behavior:
The n8n configuration workflow node must point directly to your deployed Components Service cluster instance.
vectorSearchUrlDerived using your internal cluster service path string.
const graphRagComponentsUrl="http://graphrag-components:8080";// Resolves tohttp://graphrag-components:8080/vector/searchvectorSearchUrl:'${graphRagComponentsUrl}/vector/search',vectorSearchTopKControls the exact number of nearest-match document text blocks returned per query execution loop (Default: 5)
The pipeline relies on targeted presets to parse provider-specific document payloads.
vectorStorePresetSet this to match your vector target engine instance (e.g., "
opensearch_native", "elasticsearch_native", "bedrock_knowledgebase", "pinecone", "qdrant", or "weaviate")vectorStoreCustomMappinsIf selecting the "
custom" preset flag, you must manually populate this nested object schema matching your target data model mapping.Note
Maintain this exact string spelling inside the configuration workflow node template.
vectorIndexSpecify the destination target index or collection ID (Required only when utilizing standalone engines like Pinecone, Qdrant, or Weaviate).
Each built-in preset maps standard system outputs to provider-specific document payloads.
Preset Name | Payload Field | Source Field | Score / ID Fields |
|---|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
custom | (user-defined) | (user-defined) | (user-defined) |
Important
When selecting the custompreset, you must manually populate the schema structures inside the vectorStoreCustomMappingobject parameter of the configuration node.
vectorStoreMetadataPassthroughToggles whether search metadata properties are forwarded to downstream workflows for analytical tracking or validation debugging (Default:
true).vectorStoreMetadataExcludeStrips specific sensitive fields or verbose keys (e.g.,
["content"])out of the metadata payload package prior to transit to save bandwidth.
Defines the model used by the pipeline to transform text blocks into high-dimensional numerical vectors. The Embedding Model selection is centralized entirely inside the runtime layer.
Configuration parameters must be defined inside the n8n Configurationsub-workflow block as a mandatory runtime requirement:
embeddingsProviderIdentifies the hosting framework environment (e.g.,
"aws","openai", or"custom").embeddingsModelIdThe explicit model architecture ID string utilized (e.g.,
"cohere.embed-english-v3").
Provider Reference Matrix
AWS Bedrock:
amazon.titan-embed-text-v1(1536 dimensions),amazon.titan-embed-text-v2,cohere.embed-english-v3(1024 dimensions),cohere.embed-multilingual-v3OpenAI:
text-embedding-3-small,text-embedding-3-largeCustom:
Routes to a custom, user-managed embedding models endpoint integrations
Caution
Modifying your embedding model identifier to a framework that alters dimensional boundaries requires wiping and re-embedding all historically indexed tenant documents. Ensure vector field dimensional properties match before making structural changes.