Output Parser¶

The Output Parser sub-component transforms raw LLM text output into structured data. It supports JSON extraction, regex matching, and list parsing, making it essential for nodes that need to produce machine-readable outputs from natural language responses.

Ports¶

Inputs¶

Outputs¶

Configuration¶

The Output Parser component accepts the following configuration in extra_config:

Parser types¶

JSON parser (json)¶

Extracts JSON from the LLM response. Handles two formats:

1. Direct JSON -- the entire response is valid JSON
2. Markdown code blocks -- JSON wrapped in ```json ... ``` fences

If neither format is found, the raw text is returned as-is.

Regex parser (regex)¶

Applies the configured pattern as a Python regular expression to the text and returns all matches via re.findall(). If no matches are found, the raw text is returned.

List parser (list)¶

Splits the text by newlines and strips common list prefixes (bullets, numbers, dashes). Each non-empty line becomes an element in the output array.

Input:

1. First item
2. Second item
- Third item
* Fourth item

Output:

["First item", "Second item", "Third item", "Fourth item"]

Usage¶

1. Add an Output Parser node from the Node Palette (Sub-components category)
2. Configure the parser_type and any parser-specific settings in Extra Config
3. Connect the Output Parser to an AI node via the slate diamond output_parser handle

Source resolution¶

The Output Parser reads its input text in this order:

1. If source_node is set, reads that node's output from state.node_outputs
2. Otherwise, reads the last message's content from the messages list in state

Compatible AI nodes¶

The Output Parser connects to these AI node types:

Example¶

A categorizer uses an output parser to extract the category from the LLM's JSON response:

flowchart LR
    T[Chat Trigger] --> C[Categorizer]
    M[AI Model] -.->|model| C
    OP[Output Parser] -.->|output_parser| C

    style OP fill:#94a3b8,color:white

Output Parser Extra Config:

{
  "parser_type": "json"
}

If the LLM responds with:

{"category": "technical_support", "confidence": 0.92}

The Output Parser produces {"category": "technical_support", "confidence": 0.92} as structured data that the Categorizer can use to set its category output.

Regex example¶

Extract email addresses from LLM output:

{
  "parser_type": "regex",
  "pattern": "[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,}"
}

Input: "Contact us at support@example.com or sales@example.com"

Output: ["support@example.com", "sales@example.com"]