Streams

Streams are project-scoped data stores that persist the outputs of your graph executions. Each stream has a versioned schema and supports queries with filtering, sorting, and pagination. Use streams to accumulate structured data over time and to inject stored context into downstream graph nodes.

Overview

• Project-scoped: Each stream belongs to a single project. Streams have a 1:1 relationship with graphs -- when a graph has an associated stream, exit node outputs are automatically persisted to it after each execution.
• Schema versioned: Stream schemas support draft and production versions. Define and iterate on a draft schema, then promote it to production when it is ready. This ensures that downstream consumers always read data in a known shape.
• Queryable: Query stream data with filter expressions, sort orders, limit, and offset. Stream nodes use these queries to retrieve the right slice of data at runtime.
• Toggleable: Streams can be enabled or disabled. Disabling a stream stops new data from being written without deleting existing records.

Creating and managing streams

Create and manage streams in the Portal under Storage > Streams, or through the API.

1. Create a stream and define its schema. The schema specifies the fields and types that the stream stores. If you associate the stream with a graph, the schema is derived from the output schemas of that graph's exit nodes.
2. Version the schema as your data shape evolves. Work on a draft version, then promote it to production. Previous versions remain available for reference.
3. Enable or disable the stream to control whether new execution outputs are persisted.

How data flows into streams

When a graph with an associated stream completes execution, the outputs from its exit nodes are automatically written to the stream. You do not need to configure this manually -- the association between the graph and stream handles persistence.

This makes streams well suited for accumulating results over many executions: each run appends new records that you can query later.

Querying stream data

Stream data supports the following query parameters:

You can run queries from the Portal to browse stream data, or use stream nodes in a graph to query data at execution time.

Using streams in graphs

Stream nodes read from a stream and pass the results to downstream nodes. This enables "retrieve then reason" workflows where you fetch a relevant slice of stored data and feed it into an AI or code node.

1. Add a stream node to your graph and select the target stream.
2. Configure the query with optional filter, sort, limit, and offset parameters.
3. Connect the stream node to downstream nodes. The query results become part of the input for the next step in the graph.

Set a limit to control how much context is sent to downstream nodes, and use sort to ensure the most relevant rows appear first.

Best practices

• Keep schemas stable. Use draft versions to iterate, and only promote to production when the schema is finalized. This prevents breaking changes for consumers of the stream data.
• Use filters to scope context. When injecting stream data into LLM nodes, filter for the most relevant records rather than sending everything. This improves response quality and reduces token usage.
• Set reasonable limits. Cap the number of rows returned by stream queries to avoid oversized payloads in your graph executions.
• Disable unused streams. If a graph no longer needs to persist outputs, disable the stream rather than deleting it to preserve historical data.

Related

• Stream node -- Configure stream nodes to query data within a graph.
• Graphs -- How exit node outputs are persisted to streams.
• Schemas -- Define data shapes for validation and type safety.