Manifest

Learn how to structure and write a manifest.

Writing a Manifest

A manifest in your site defines the execution flow of tools and scripts, specifying dependencies and data transformations. This document outlines how to structure and write a manifest following the format used in your system.

Writing Your Own Manifest

To create a manifest:

1. Define an id and name for each step.
2. Choose the appropriate tool, such as INPUT_JSON or PYTHON_SANDBOX.
3. Specify input values under input.
4. List dependencies using needs to ensure correct execution order.

By following this structure, you can build a clear, modular workflow for processing data efficiently.

Step Components

Each step in a Jinbaflow manifest is composed of several key components that define its behavior and relationship with other steps. Here’s a detailed breakdown:

1. Basic Step Structure

- id: step_id                 # Unique identifier for the step
  tool: TOOL_NAME            # The tool to use (e.g., LLM, PYTHON_SANDBOX)
  input:                     # Input parameters required by the tool
    - name: parameter_name   # Name of the input parameter
      value: parameter_value # Value for the parameter
  config:                    # Optional: Tool-specific configuration
    - name: config_name     # Name of the configuration parameter
      value: config_value   # Value for the configuration
  needs:                     # Optional: List of step IDs that must complete first
    - step_id1             # Step that must complete before this one
    - step_id2             # Another dependency
  when: "condition"          # Optional: Condition that must be true for step to run
  forEach: "items"           # Optional: Run step for each item in a collection

Each field serves a specific purpose:

• id: Required unique identifier for referencing the step
• tool: Required name of the tool to execute (must be one of the available tools)
• input: Required list of input parameters specific to the tool
• config: Optional tool-specific configuration (e.g., API tokens, temperature for LLM)
• needs: Optional list of step IDs that must complete before this step runs
• when: Optional condition that must evaluate to true for the step to execute
• forEach: Optional field to execute the step multiple times over a collection

2. Configuration

Used to set tool-specific settings:

config:
  - name: config_name
    value: config_value

Example:

config:
  - name: temperature
    value: 0.7
  - name: token
    value: "{{secrets.API_TOKEN}}"

3. Dependencies

Specify which steps must complete before this step can run:

needs:
  - step_id1
  - step_id2

Example:

- id: summarize
  tool: LLM
  needs:
    - fetch_content
    - analyze_sentiment

4. Conditional Execution

Control whether a step should run based on conditions:

when: "<condition_expression>"

Examples:

when: "'{{steps.check_genre.result.content}}' == 'General'"
when: "{{steps.count_items.result}} > 5"

5. Loop Execution

Execute a step for each item in a collection:

forEach: "{{steps.input_list.result}}"

Example:

- id: process_items
  tool: PYTHON_SANDBOX_RUN
  forEach: "{{steps.get_items.result}}"
  input:
    - name: code
      value: "process_item('{{item}}')"
    - name: data_type
      value: "STRING"

6. Result Reference

Reference results from previous steps:

value: "{{steps.step_id.result}}"           # Access entire result
value: "{{steps.step_id.result.content}}"   # Access specific property

Important Notes

• Step IDs: Must be unique within the workflow
• Dependencies:
  • Create an implicit execution order
  • Can form complex graphs (including diamond patterns)
  • Circular dependencies are not allowed
• Conditions:
  • Support basic comparisons and logical operators
  • Can reference results from previous steps
  • Evaluated before step execution
• Variable References:
  • Use {{steps.step_id.result}} to access previous step results
  • Use {{secrets.SECRET_NAME}} to access configured secrets
  • Use {{item}} in forEach loops to access the current item
• YAML Syntax:
  • Use | for a multi-line string until the indentation ends.

Jinja2 Template Syntax

Jinbaflow supports Jinja2 templating for dynamic value generation and complex logic in manifest files. This powerful feature allows you to create more flexible and reusable workflows. When using templates, you have access to:

• steps: Results from previous steps
• item: Current item in forEach loops
• secrets: Configured secrets
• Built-in functions like zip(), range(), etc.

Use {%- and -%} to control whitespace in the output, where the - removes whitespace before or after the block.

Basic Syntax

1. Variable Interpolation

value: "{{ variable }}"

2. Control Structures

# For loops
{%- for item in items %}
  {{ item }}
{%- endfor %}

# Conditional statements
{%- if condition %}
  {{ value }}
{%- endif %}

3. Accessing Step Results

value: "{{ steps.step_id.result }}"
value: "{{ steps.step_id.result.content }}"

Common Use Cases

1. Processing Lists

- id: process_items
  tool: PYTHON_SANDBOX_RUN
  input:
    - name: code
      value: |-
        items = []
        {%- for item in steps.input_list.result %}
        items.append({
          "id": "{{item.id}}",
          "value": "{{item.value}}"
        })
        {%- endfor %}
    - name: data_type
      value: "STRING"

2. Conditional Logic

- id: conditional_step
  tool: PYTHON_SANDBOX_RUN
  input:
    - name: code
      value: |-
        {%- if steps.check.result.status == 'success' %}
        print("Processing successful case")
        {%- else %}
        print("Handling error case")
        {%- endif %}
    - name: data_type
      value: "STRING"

3. Combining Multiple Lists

- id: combine_data
  tool: PYTHON_SANDBOX_RUN
  input:
    - name: code
      value: |-
        results = []
        {%- for item1, item2 in zip(list1, list2) %}
        results.append(f"{item1}: {item2}")
        {%- endfor %}
    - name: data_type
      value: "STRING"

Common Patterns

1. Filtering Lists

{%- for item in items if item.status == 'active' %}
{{ item.name }}
{%- endfor %}

2. Setting Default Values

{{ variable | default('default_value') }}

3. String Manipulation

{{ variable | upper }}
{{ variable | trim }}
{{ variable | replace('old', 'new') }}

Example Manifest

Below is an example manifest that processes numerical data through multiple steps:

{- id: input_data
  name: input_data
  tool: INPUT_JSON
  config: []
  input:
    - name: value
      value: "[1, 2, 3, 4, 5]"

- id: branch_a1
  name: branch_a1
  tool: PYTHON_SANDBOX
  config: []
  input:
    - name: code
      value: |
        data = {{steps.input_data.result}}
        result = sum([x for x in data if x % 2 == 0])
        result  # even numbers sum
  needs:
    - input_data

- id: branch_a2
  name: branch_a2
  tool: PYTHON_SANDBOX
  config: []
  input:
    - name: code
      value: |
        result = {{steps.branch_a1.result.text}} * 2
        result  # double the sum
  needs:
    - branch_a1}

Breakdown of Each Step

1. Input Data (input_data)

• Uses the INPUT_JSON tool to define an array of numbers [1, 2, 3, 4, 5].

2. Branch A1 (branch_a1)

• Uses PYTHON_SANDBOX to compute the sum of even numbers from input_data.
• Depends on input_data.

3. Branch A2 (branch_a2)

• Doubles the result from branch_a1.
• Depends on branch_a1.

4. Branch B1 (branch_b1)

• Computes the sum of odd numbers from input_data.
• Depends on input_data.