Non-JSON Mode

Use json_mode: false when working with models that cannot produce structured JSON reliably — local models via Ollama, small open-source models, or any provider where JSON formatting is unreliable.

How It Works

In JSON mode (the default), each action uses a schema to define its output fields, and the framework injects the schema into the prompt so the LLM returns structured JSON.

In non-JSON mode, schemas are not injected. Instead, each action uses output_field to name a single field. The LLM returns plain text, and the framework maps the raw response to that field name.

	JSON Mode (true)	Non-JSON Mode (false)
Output definition	schema: my_schema	output_field: field_name
LLM returns	Structured JSON matching schema	Plain text
Validation	Schema validation + reprompt	None (raw text stored)
Model requirements	Must support JSON output	Any model works
Fields per action	Multiple	One

Field-by-Field Construction

The key pattern: instead of one action producing a complex JSON object with many fields, you chain multiple actions where each produces one field. By the end of the pipeline, all fields accumulate into a complete record.

defaults:
  json_mode: false
  model_vendor: ollama
  model_name: llama3

actions:
  - name: classify_topic
    output_field: topic              # LLM response → stored as "topic"
    prompt: $my_workflow.Classify

  - name: assess_sentiment
    dependencies: [classify_topic]
    output_field: sentiment          # LLM response → stored as "sentiment"
    prompt: $my_workflow.Sentiment

  - name: write_summary
    dependencies: [classify_topic, assess_sentiment]
    output_field: summary            # LLM response → stored as "summary"
    prompt: $my_workflow.Summary

After all three actions, the record contains {topic, sentiment, summary} — structured output assembled from plain text responses.

Prompt Style

Non-JSON prompts should ask for one answer in the simplest possible format:

{prompt Classify}
**Title**: {{ source.title }}
**Body**: {{ source.body }}

Classify this into ONE of: bug, feature_request, question, account_issue.

Answer with the category name only. Nothing else.
{end_prompt}

The "answer with X only, nothing else" instruction is important. Without it, smaller models tend to add explanations or formatting that pollute the field value.

Per-Action Model Override

Non-JSON mode works with per-action model overrides. Use a cheap local model for simple classification, and override to a stronger model for actions that need better reasoning:

defaults:
  json_mode: false
  model_vendor: ollama
  model_name: llama3                 # Cheap: local, no API cost

actions:
  - name: classify_issue
    output_field: issue_type
    prompt: $workflow.Classify        # Simple classification → local model is fine

  - name: draft_response
    output_field: suggested_response
    prompt: $workflow.Draft_Response
    model_vendor: openai              # Override: customer-facing text needs quality
    model_name: gpt-4o-mini
    api_key: OPENAI_API_KEY

Guards Still Work

Guards evaluate upstream field values the same way regardless of JSON mode:

  - name: draft_response
    output_field: suggested_response
    guard:
      condition: 'severity != "low"'  # Skip for low-severity → save tokens
      on_false: "skip"

Context Scope Still Works

Progressive context narrowing applies identically. Later actions can observe specific upstream fields:

  - name: draft_response
    context_scope:
      observe:
        - classify_issue.issue_type    # See the classified type
        - summarize_issue.summary      # See the summary
      drop:
        - source.body                  # Don't pass raw body → smaller prompt

When to Use Non-JSON Mode

Scenario	Recommendation
Ollama / local models	Use non-JSON mode
Small models (< 7B params)	Use non-JSON mode
Models without JSON support	Use non-JSON mode
Cost-sensitive pipelines	Mix: non-JSON for simple steps, JSON for complex
Prototyping / testing	Non-JSON is faster to set up (no schemas)
Production with capable models	JSON mode with schemas for validation