Milvus. The selected vector store provider for this configuration.

Required. Endpoint for your Milvus instance (for example http://localhost:19530 for a local default). For Milvus Cloud or Zilliz Cloud, use the connection URI from your cluster or project details page.

Required. Length of each embedding vector stored in this collection. Must match your embedding model’s output dimension. Common examples include 1536 (for example OpenAI text-embedding-3-small at default size) and 384 (for example All-MiniLM-style models)—use the value your embedding pipeline actually produces.

Optional with a product default (often default). Logical collection that holds vectors and schema. If the collection does not exist, the product may create it using this dimension and index settings—confirm auto-create behavior in your release notes.

Optional with a product default (often default). Milvus database name for namespace isolation. If omitted, the integration typically uses the default database.

Distance metric for similarity search (for example Cosine, L2 (Euclidean), IP (inner product)). Choose a metric that matches how your embeddings were trained or normalized; cosine is common for normalized embedding vectors.

Vector index algorithm. Typical options include HNSW (strong query performance, higher memory), IVF_FLAT (balance of speed and memory), IVF_SQ8 (more memory-efficient), FLAT (exact search; slower at large scale), and AUTOINDEX (automatic choice where supported). Pick based on dataset size, latency budget, and memory.

Optional for local Milvus without auth. For Milvus Cloud or Zilliz Cloud, usually required. Leave blank only when your deployment does not use token authentication.

Optional. User name for basic authentication when the Milvus endpoint requires user/password auth. Not used when you authenticate only with an API key.

Optional. Password paired with username for basic authentication. Treat stored passwords as secrets.

Optional. Maximum number of hits to return from a similarity search. Must be a positive integer when set. If empty, the product uses its default top K.

Optional. Minimum similarity score for a hit to be returned; results below the threshold are dropped. For many metrics, meaningful ranges are between 0 and 1; leave empty to return all top-K results without a score cutoff.

Maximum time to wait for a single RPC or request to finish. Default is often 60000 ms (60 seconds).

Maximum time to wait while establishing a connection. Default is often 20000 ms (20 seconds).

When enabled, keeps connections warm (for example HTTP/2 keep-alive), which can reduce latency on steady workloads. Default is often on.

Interval between keep-alive pings when the connection is idle. Default is often 30000 ms (30 seconds).

How long to wait for a keep-alive ping acknowledgment. Default is often 5000 ms (5 seconds).

How long a connection may stay idle before the client closes it. Default is often 600000 ms (10 minutes).

Per-call gRPC deadline. When set above 0, some clients enable waitForReady so calls queue on broken channels instead of failing immediately. 0 usually means disabled. Default is often 0.

How many times to retry after a failed request. Default is often 3.

When enabled, retries when the server signals rate limiting. Default is often on.

Delay before the first retry. Default is often 200 ms.

Upper cap on delay between retries. Default is often 1000 ms (1 second).

Factor for exponential backoff between attempts. Default is often 2.

None — skip embedding configuration for this flow (only if valid for your use case). Pre-configured — use an embedding profile managed under AI Services > Embedding models. Custom — define or override embedding settings inline in this screen.

After you choose a source, pick the embedding provider (OpenAI, Mistral, Ollama, and so on). That choice drives which embedding API runs; Milvus settings above define where those vectors are stored and queried.