2.1.18 Frontend R2R

R2R

Handle: r2r
Dashboard: http://localhost:34273
API: http://localhost:34272

R2R provides document ingestion, semantic search, and retrieval-augmented answers. Harbor uses upstream's lighter deployment: R2R API and dashboard, a persistent pgvector database, and the graph-clustering service. It does not include the full deployment's Hatchet, Unstructured, or object-storage stack, so advanced extraction and orchestration features may require that upstream deployment instead.

Starting

Set three unique secrets. The preflight container rejects empty, short, or non-alphanumeric secrets. harbor config set normally prints values, so redirect its output when setting secrets. Save the three values securely along with database backups; changing them can break access to existing accounts or data.

harbor config set r2r.db_password "$(openssl rand -hex 32)" >/dev/null 2>&1
harbor config set r2r.secret_key "$(openssl rand -hex 32)" >/dev/null 2>&1
harbor config set r2r.admin_password "$(openssl rand -hex 32)" >/dev/null 2>&1
harbor up r2r ollama --open

Sign in to the dashboard as admin@harbor.local using the generated r2r.admin_password. The dashboard's deployment URL should show http://localhost:34272. Starting with ollama selects Harbor's local-model profile, pulls the configured chat and embedding models, and points R2R at Harbor Ollama. To ingest a local text document, use the dashboard's Documents page or the API with ingestion_mode=fast; then use Search or Chat. The default chat model is qwen2.5:3b, and the embedding model is mxbai-embed-large (1,024 dimensions). The local profile disables automatic extraction. Image/audio parsing, advanced extraction, and graph workflows are not covered by this local path and may need additional providers or the upstream full deployment.

Starting harbor up r2r without ollama still loads the dashboard and existing data, but ingestion, semantic search, and RAG need ollama running. Harbor keeps the same 1,024-dimensional embedding schema in both startup modes so switching between them does not invalidate the database. Other provider configurations require a custom R2R config and are not part of this service's default setup. Never commit provider keys to the repository.

Configuration and persistence

Setting	Default	Purpose
r2r.dashboard_port	34273	Browser dashboard port
r2r.api_port	34272	Browser and SDK API port
r2r.bind_host	127.0.0.1	Host address for both published ports
r2r.public_host	localhost	API hostname advertised to the dashboard
r2r.db_password	Empty	Required database password
r2r.secret_key	Empty	Required application cryptographic secret
r2r.admin_password	Empty	Required bootstrap admin password
r2r.chat_model	qwen2.5:3b	Ollama chat model for RAG responses
r2r.embedding_model	mxbai-embed-large	Ollama embedding model; must output 1,024 dimensions

Both ports are loopback-only by default. For a trusted LAN browser, set both r2r.bind_host (for example 0.0.0.0) and r2r.public_host to the address used by the browser, then restart. Do not expose the API or dashboard to the public internet without a trusted TLS/authentication proxy. The r2r-db-data Docker volume persists documents, embeddings, and accounts through harbor down r2r; services/r2r/user_tools/ is a host workspace for optional custom tools. Back up the volume, this directory if used, and the three secrets together.

Troubleshooting

• curl http://localhost:34272/v3/health should return {"results":{"message":"ok"}}. If the dashboard cannot connect, check its deployment URL and reachability of the API port from the browser.
• For local text ingestion, start both r2r and ollama. Use fast ingestion to avoid extraction services not included in this deployment.
• If RAG returns an empty answer with a reasoning model, switch r2r.chat_model to a non-reasoning model such as the default qwen2.5:3b, then restart.
• After changing a distributed default in a Git checkout, run harbor config update to propagate it to the local .env.

See the upstream Docker deployment for full-stack options.