Available starting with FlowX.AI 5.9.1.
ai-platform-kb-enrichment replaces the decommissioned di-platform service.Overview
ai-platform-kb-enrichment provides design-time document intelligence over knowledge-base documents — chunking, enrichment, and preparing content for indexing. It is a Python service in the AI Platform, exposing gRPC and consuming a Kafka enrichment-request topic.
It deploys as a sub-chart of the AI Platform umbrella Helm chart like every other AI service (listening on port 9100 via SERVICE_PORT in Kubernetes), but has a few extra deployment requirements covered below.
Prerequisites
Infrastructure
- AI Platform chart deployed (KB Enrichment is a subchart)
- PostgreSQL instance (shared with the other AI Platform databases)
- S3-compatible object storage (MinIO, AWS S3, …)
- Kafka
Platform dependencies
doc-parserdeployed (doc-parser.installed: true)- An organization
TEXT_GENERATIONLLM capability configured - Egress to
huggingface.coon first start (or a pre-seeded model cache)
Deployment
KB Enrichment is enabled as part of the AI Platform chart. The dependencies below are provisioned automatically by the FlowX Helm charts in standard deployments — the notes call out what each one needs.PostgreSQL database
KB Enrichment uses a dedicatedkbenrichment database on the same instance and credentials as the other AI Platform databases (for example, next to the embedder database). The FlowX Helm dependencies chart provisions it automatically.
KB Enrichment uses PostgreSQL only for service/job state — vector embeddings are stored in Qdrant. The
knowledgebases_design Qdrant collection is auto-created by the embedder service; no provisioning step is required.Object storage
Parse offloads go to thekb-enrichment-bucket bucket (auto-created on MinIO), shared with Knowledge Base Indexer v2: the indexer writes offload documents and KB Enrichment reads them back.
| Environment Variable | Description | Default Value |
|---|---|---|
KB_ENRICHMENT_OFFLOAD_BUCKET | Optional override for the offload bucket name. Leave unset to use the shared default. | kb-enrichment-bucket |
Model cache and egress
On first start the service downloads itsmodel2vec chunking model from huggingface.co. The Helm chart already points the model caches at writable paths, so no manual configuration is needed:
| Environment Variable | Description | Default Value |
|---|---|---|
HF_HOME | Hugging Face cache directory | /tmp |
XDG_CACHE_HOME | XDG cache directory | /tmp |
NLTK_DATA | NLTK data directory | /tmp |
Kafka topic
KB Enrichment consumes enrichment requests produced by Knowledge Base Indexer v2:| Topic | Partitions | Direction |
|---|---|---|
ai.flowx.ai-platform.knowledgebase.internal.enrichment-request.v1 | 3 | Indexer v2 → KB Enrichment (consume) |
Required platform configuration
Turn on doc-parser
AI Base ships with
doc-parser.installed: false. Document ingestion through KB Enrichment requires doc-parser to be deployed — set doc-parser.installed: true in your AI Platform values. See the Document Parser setup guide.Seed the LLM configuration
The organization needs a
TEXT_GENERATION capability with a default model and provider API key configured in the AI Providers UI. Without it, enrichment fails with a no capability TEXT_GENERATION error. See AI providers.Service accounts and Keycloak clients follow the standard AI Platform pattern — nothing new is required for KB Enrichment.
Verification
Confirm dependencies
Verify the
kbenrichment database exists, the kb-enrichment-bucket was created on object storage, and the enrichment-request topic is present.Related resources
AI Platform setup
Full AI Platform deployment and shared configuration
Document Parser setup
Required dependency for document ingestion
AI providers
Configure the TEXT_GENERATION LLM capability
Qdrant setup
Vector store for knowledge-base embeddings