← Docs Index

Short answers with links to the full detail. New here? Read Meet Intel® AI for Enterprise RAG first.

Getting started

What is Intel® AI for Enterprise RAG?

A complete retrieval-augmented generation application you deploy on your own hardware: document ingestion, vector search, retrieval pipelines, guardrails, chat UI, single sign-on, and dashboards. It installs as the erag layer of Intel® AI for Enterprise Solutions. See Meet.

Which repository do I clone?

Clone github.com/intel/enterprise-ai-solutions and run every command from its root. init erag clones this repository into ext/enterprise.ai-erag at a pinned revision. You never clone this repository directly to deploy it.

Do I have to install the platform before the RAG layer?

No. install erag pulls its dependencies automatically, in order: infrastructure, platform, inference, then the RAG layer. Installing the platform first is optional and useful only when you want to confirm model serving works before adding RAG. See Quick Start.

How long does a deployment take?

The RAG layer itself takes 20 to 30 minutes once the platform is up. A full deployment from bare metal, including Kubernetes provisioning and model downloads, takes longer and depends on your network.

What hardware do I need?

On top of the platform: 60 logical cores, 128 GB RAM, and 200 GB disk for a multi-user workload. A limited single-user configuration runs on 32 logical cores and 64 GB RAM for evaluation only, and requires tuning. See Prerequisites.

Do I need a GPU or an accelerator?

No. Model serving is tuned for Intel® Xeon® CPUs, including NUMA-aware CPU pinning.

Do I need internet access?

Yes, for container images and model weights. Proxy settings are configured in the platform's global_config.yaml. To deploy from your own registry instead, see Building Images.

Do I need a Hugging Face token?

Only for gated models such as Llama, Mistral, or Gemma. Export HF_TOKEN before installing. Open models need nothing.

Pipelines and configuration

What is the difference between a flavour, a pipeline type, and a variant?

A flavour is the preset that init erag --flavour <name> seeds your configuration from. A pipeline type is the service graph that gets composed (chatqna, docsum, translation). A variant modifies that graph, for example inserting a query-rewrite step. See Pipelines.

Which flavours are available?

chatqna (default, conversational retrieval), docsum (document summarization), translation, audioqna (voice), and pl_chatqna (Polish-language retrieval).

Can I run more than one pipeline at the same time?

No. One pipeline is deployed at a time. Switching re-composes the service graph and tears down the previous pipeline; the vector store lives in its own namespace and is preserved. See Pipelines.

How do I switch pipelines after installing?

Edit pipeline_type and pipeline_variant in env/<name>/config.erag.yaml, then re-run install erag. You do not re-run init.

Why does init erag --flavour refuse to run on an environment I already created?

To protect edits you have already made to the configuration. Either use a fresh --env, or delete env/<name>/config.erag.yaml to reseed it from the flavour.

Which vector databases are supported?

redis-cluster (default) and mssql, selected with vector_databases_vector_store. See Configuration.

How do I change the LLM, embedding, or reranking model?

List the models under inference_models in env/<name>/config.erag.yaml. Each entry must exist in env/<name>/models-rag.yaml. See Models.

Do I need to deploy models manually with model-manager?

No. install erag deploys the models declared in inference_models through its own step, sizes them, and undeploys models that are no longer needed when you switch pipelines.

What happens if I change the embedding model after ingesting documents?

Embedding dimensions change and the existing index is not rebuilt automatically, so retrieval will return nothing useful. Re-ingest your documents after the change. See Pipelines - best practices.

Where do my settings live, and why do nested keys not work?

Everything is under env/<name>/: config.erag.yaml for RAG settings, global_config.yaml for platform settings, models-rag.yaml for the model catalog. The override surface is flat because files are merged shallowly; a nested dictionary replaces the role default instead of merging into it. See Configuration.

Can I keep several deployments side by side?

Yes. Each --env <name> has its own configuration, credentials, and kubeconfig. All environments share one ext/ checkout, so they cannot be pinned to different revisions of this repository.

Access, identity, and security

Where are the credentials after installing?

In env/<name>/logs/rag/default_credentials.yaml. They are one-time passwords; change them after first login. A few service credentials live in Kubernetes secrets instead, listed in Install.

How do I open the UI?

The gateway binds ports 80 and 443 on the cluster node. Add the base domain and its subdomains to /etc/hosts on the machine with the browser, forwarding with ssh -L 443:localhost:443 if that machine is not the cluster node. See Install - access the UI.

Why does the browser warn about the certificate?

Self-signed certificates are the default. Visit https://s3.<base_domain_name> and accept the warning before ingesting documents, otherwise uploads fail silently in the browser.

Can I use our corporate identity provider?

Yes. Keycloak federates to Microsoft Entra ID, with multi-factor authentication and Active Directory federation. See Authentication. The same Entra ID app registration also enables SharePoint ingestion.

Can it ingest from SharePoint or an object store?

Yes. SharePoint Online with scheduled sync is covered in SharePoint; S3-compatible and NetApp ONTAP S3 backends are covered in Object Store.

Is traffic between services encrypted?

RAG namespaces join the platform's Istio ambient mesh, which provides mutual TLS without sidecars. The current posture and its limits are documented in Mesh.

Can AI agents use the retrieval and ingestion capabilities?

Yes, through the Model Context Protocol gateway. It is off by default; set mcp_enabled: true. See MCP.

Running it

How do I check that the deployment is healthy?

./es_auto_installer.sh status --env <name> for namespaces, pods, and endpoints, and ./es_auto_installer.sh validate erag --env <name> for health checks. The pipeline test scripts under ext/enterprise.ai-erag/deployment/scripts/ send a real request.

Where are the logs?

Installer and Ansible logs are in env/<name>/logs/, for example install-erag-*.log. Service logs are in Loki and browsable from Grafana. See Telemetry.

What can I see in Grafana?

RAG-specific dashboards for per-service traffic, the ingestion pipeline, autoscaling, and LLM inference, alongside the platform dashboards. See Telemetry.

An install failed halfway. Can I just re-run it?

Yes. Installs are idempotent, so re-running continues rather than duplicating work. Check the log in env/<name>/logs/ first to fix the underlying cause.

How does it scale under load?

Horizontal Pod Autoscaling is enabled by default (hpa_enabled: true) and disabled for minimum-hardware deployments. Per-step resources, replica sizing, and vector-database tuning are in Performance.

How do I remove it?

teardown erag removes the RAG layer in reverse order and undeploys its models, leaving the cluster, platform, and inference layers running. teardown infrastructure removes the cluster and everything on it. Teardown is environment-scoped: use the same --env you installed with.

Will switching models or pipelines re-download the weights?

No. Weights stay on the model-store persistent volume, so a model that comes back is re-placed rather than re-downloaded.

Can I use my own container registry?

Yes. Build the component images and point registry and tag in config.erag.yaml at your registry. See Building Images.

Something is broken and none of this covers it.

Start with Troubleshooting, which maps symptoms to causes and includes the debug collection tool. If you need further support, access supporttickets.intel.com