Workflow Functions

Workflow functions allow you to use Otoroshi workflows as tool function backends. When the LLM decides to call a tool, the gateway executes the referenced workflow with the tool arguments as input and returns the workflow result to the LLM.

This is the most powerful backend kind — workflows can chain HTTP calls, transform data, apply conditional logic, call other LLM providers, query vector stores, and much more.

How it works

1. The LLM decides to call the tool with structured arguments (e.g., {"query": "latest news"})
2. The gateway looks up the referenced Otoroshi workflow by ID
3. The tool arguments are passed as the workflow input (${workflow_input})
4. The workflow executes its steps (HTTP calls, data transformations, LLM calls, etc.)
5. The workflow returned value is sent back to the LLM as the tool result

The models a workflow calls are accounted for on their own: each one is billed, audited and counted against its budgets as the model it is. The chat call that ran the tool keeps the cost and the footprint of its own model.

Configuration

{
  "name": "my_workflow_tool",
  "description": "A tool backed by an Otoroshi workflow",
  "strict": true,
  "parameters": {
    "query": {
      "type": "string",
      "description": "The search query"
    }
  },
  "required": ["query"],
  "backend": {
    "kind": "Workflow",
    "options": {
      "workflow_id": "workflow-entity-id"
    }
  }
}

Options

Parameter	Type	Description
workflow_id	string	The ID of the Otoroshi workflow entity to execute

Input handling

The tool arguments are passed as the workflow input and accessible via ${workflow_input} in the workflow steps:

• If the arguments are a JSON object (e.g., {"city": "Paris", "unit": "celsius"}), each key is directly accessible: ${workflow_input.city}, ${workflow_input.unit}
• If the arguments are a plain string, they are wrapped as {"input": "the string"} and accessible via ${workflow_input.input}

Output handling

The workflow's returned value is serialized to JSON and sent back as the tool result. If the workflow terminates with an error, the error is returned as {"error": ...}.

What are Otoroshi workflows?

Otoroshi workflows are a JSON-based orchestration engine built into Otoroshi. They allow you to build complex automation pipelines using a visual editor or JSON configuration.

A workflow is a directed graph of nodes that execute sequentially or in parallel. Key node types include:

Node	Purpose
workflow	Sequential step execution
call	Invoke a function (HTTP, LLM, store, etc.)
assign	Set variables in workflow memory
if / switch	Conditional branching
foreach / map / filter	Array iteration and transformation
parallel	Concurrent execution
try	Error handling
value	Return a literal value

Built-in functions

Workflows come with 30+ built-in functions including:

• HTTP: core.http_client — make HTTP requests to external APIs
• Storage: core.store_get, core.store_set, core.store_del — key-value storage
• System: core.system_call, core.env_get, core.config_read
• Workflow: core.workflow_call — call other workflows
• WASM: core.wasm_call — execute WASM plugins

The LLM extension registers additional workflow functions for AI operations (LLM calls, embeddings, vector store, moderation, etc.).

Basic workflow structure

{
  "kind": "workflow",
  "steps": [
    {
      "kind": "call",
      "function": "core.http_client",
      "args": {
        "url": "https://api.example.com/search?q=${workflow_input.query}",
        "method": "GET"
      },
      "result": "api_response"
    },
    {
      "kind": "assign",
      "values": {
        "formatted_result": {
          "$jq": {
            "value": { "$mem_ref": { "name": "api_response" } },
            "filter": ".body | fromjson | .results"
          }
        }
      }
    }
  ],
  "returned": { "$mem_ref": { "name": "formatted_result" } }
}

For full documentation on workflows, see the Otoroshi workflows documentation.

Examples

RAG search tool

A tool that searches a vector store and returns relevant context:

{
  "name": "search_knowledge_base",
  "description": "Search the knowledge base for relevant information",
  "strict": true,
  "parameters": {
    "query": {
      "type": "string",
      "description": "The search query"
    }
  },
  "required": ["query"],
  "backend": {
    "kind": "Workflow",
    "options": {
      "workflow_id": "workflow_rag_search"
    }
  }
}

With a workflow that computes embeddings and queries a vector store:

{
  "kind": "workflow",
  "steps": [
    {
      "kind": "call",
      "function": "extensions.com.cloud-apim.llm-extension.compute_embedding",
      "args": {
        "provider": "embedding-provider-id",
        "input": "${workflow_input.query}"
      },
      "result": "embedding"
    },
    {
      "kind": "call",
      "function": "extensions.com.cloud-apim.llm-extension.vector_store_search",
      "args": {
        "provider": "vector-store-id",
        "vector": { "$mem_ref": { "name": "embedding" } },
        "limit": 5
      },
      "result": "search_results"
    }
  ],
  "returned": { "$mem_ref": { "name": "search_results" } }
}

API aggregation tool

A tool that calls multiple APIs and aggregates the results:

{
  "name": "get_city_info",
  "description": "Get comprehensive information about a city including weather and population",
  "strict": true,
  "parameters": {
    "city": {
      "type": "string",
      "description": "The city name"
    }
  },
  "required": ["city"],
  "backend": {
    "kind": "Workflow",
    "options": {
      "workflow_id": "workflow_city_info"
    }
  }
}

With a workflow that makes parallel HTTP calls:

{
  "kind": "workflow",
  "steps": [
    {
      "kind": "parallel",
      "parallels": [
        {
          "kind": "call",
          "function": "core.http_client",
          "args": {
            "url": "https://api.weather.example.com/v1/weather?city=${workflow_input.city}",
            "method": "GET"
          },
          "result": "weather"
        },
        {
          "kind": "call",
          "function": "core.http_client",
          "args": {
            "url": "https://api.geodata.example.com/v1/city?name=${workflow_input.city}",
            "method": "GET"
          },
          "result": "geo_data"
        }
      ]
    },
    {
      "kind": "assign",
      "values": {
        "result": {
          "weather": { "$mem_ref": { "name": "weather" } },
          "geo_data": { "$mem_ref": { "name": "geo_data" } }
        }
      }
    }
  ],
  "returned": { "$mem_ref": { "name": "result" } }
}

LLM sub-call tool

A tool that calls another LLM provider to process data:

{
  "name": "summarize_text",
  "description": "Summarize a long text into a concise summary",
  "strict": true,
  "parameters": {
    "text": {
      "type": "string",
      "description": "The text to summarize"
    }
  },
  "required": ["text"],
  "backend": {
    "kind": "Workflow",
    "options": {
      "workflow_id": "workflow_summarize"
    }
  }
}

With a workflow that calls an LLM provider:

{
  "kind": "workflow",
  "steps": [
    {
      "kind": "call",
      "function": "extensions.com.cloud-apim.llm-extension.llm_call",
      "args": {
        "provider": "llm-provider-id",
        "prompt": "Summarize the following text in 3 sentences:\n\n${workflow_input.text}"
      },
      "result": "summary"
    }
  ],
  "returned": { "$mem_ref": { "name": "summary" } }
}

Provider configuration

Reference workflow-backed tool functions in the provider's wasm_tools array like any other tool function:

{
  "provider": "openai",
  "connection": {
    "base_url": "https://api.openai.com/v1",
    "token": "${vault://local/openai-token}",
    "timeout": 30000
  },
  "options": {
    "model": "gpt-4o",
    "wasm_tools": [
      "tool-function_rag_search",
      "tool-function_city_info"
    ],
    "max_function_calls": 10
  }
}

Comparison with other backends

Aspect	QuickJs	HTTP	Workflow
Complexity	Simple logic	Single API call	Multi-step orchestration
Capabilities	JavaScript + HTTP	HTTP request/response	HTTP, LLM, vector store, conditions, loops, parallel execution
Data transformation	JavaScript code	response_path extraction	Full operator library ($jq, $map_get, $projection, etc.)
Error handling	try/catch in JS	HTTP error codes	try nodes with error handling steps
Best for	Quick data processing	Calling a single external API	Complex pipelines, RAG, multi-API aggregation