ffai-workflow-adapters 0.1.0

External workflow adapters for FFAI — load workflows from Airtable, Google Sheets, and other tabular sources

ffai-workflow-adapters

External workflow adapters for ffai — define and execute LLM workflows from Airtable, Excel, and other tabular sources. Each row in a spreadsheet becomes a workflow step (name, prompt, model, temperature, etc.).

Installation

pip install ffai-workflow-adapters

# With Airtable support
pip install ffai-workflow-adapters[airtable]

# With Excel support
pip install ffai-workflow-adapters[excel]

# With both
pip install ffai-workflow-adapters[all]

Quick Start

import asyncio
import os

from dotenv import load_dotenv
from ffai import FFAI
from ffai.Clients.AsyncFFLiteLLMClient import AsyncFFLiteLLMClient
from ffai_workflow_adapters import load_workflow_airtable, write_workflow_results

load_dotenv()


async def main():
    from ffai_workflow_adapters import get_config

    config = get_config()
    client_cfg = config.clients.get_client_type(config.clients.default_client)
    client = AsyncFFLiteLLMClient(
        model_string=f"{client_cfg.provider_prefix}{client_cfg.default_model}",
        api_key=os.environ.get(client_cfg.api_key_env, ""),
    )
    ffai = FFAI(client)

    base_id = os.environ["AIRTABLE_BASE_ID"]
    spec = load_workflow_airtable(base_id, "Workflow Steps", view="basic", name="my_workflow")

    result = await ffai.execute_workflow(spec)

    write_workflow_results(base_id, "_results", result)


asyncio.run(main())

How It Works

1. Define your workflow as rows in Airtable or Excel — each row is a step with a name, prompt, model, and optional parameters
2. Load the table into a WorkflowSpec using the adapter for your data source
3. Execute with ffai — steps run sequentially, with {{step.response}} interpolation chaining outputs between steps
4. Write back results (response, tokens, cost, duration) to your data source

Example Workflow Table

name	prompt	client	history	temperature
topic	Name a famous scientific discovery and explain it in one sentence.	litellm-mistral-small		0.7
explain	Given this discovery: {{topic.response}} — write a paragraph about its impact.	litellm-gpt-4o-mini	topic	0.5

Adapters

Adapter	Install	Documentation
Airtable	pip install ffai-workflow-adapters[airtable]	Airtable adapter guide
Excel	pip install ffai-workflow-adapters[excel]	Excel adapter guide

Configuration

Uses pydantic-settings with YAML files and environment variable overrides.

Config Files

File	Purpose
config/main.yaml	Retry and resilience settings
config/adapters.yaml	Per-adapter field maps, passthrough columns, named variants
config/clients.yaml	LLM client definitions (LiteLLM providers)
config/logging.yaml	Logging configuration

Priority (highest to lowest)

1. Explicit constructor kwargs
2. Environment variables (nested delimiter __, e.g., RETRY__MAX_ATTEMPTS=5)
3. Merged YAML files from config/

Clients

Define named clients in config/clients.yaml and reference them by name in your data source's client column:

default_client: litellm-mistral-small

client_types:
  litellm-mistral-small:
    type: litellm
    api_key_env: MISTRAL_API_KEY
    provider_prefix: "mistral/"
    default_model: mistral-small-latest

  litellm-gpt-4o-mini:
    type: litellm
    api_key_env: OPENAI_API_KEY
    provider_prefix: "openai/"
    default_model: gpt-4o-mini
    fallbacks:
      - mistral/mistral-small-latest

Resilience

Airtable operations are protected by rate limiting, circuit breakers, and retries with exponential backoff. All settings are in config/main.yaml and tunable via environment variables:

resilience:
  rate_limit:
    requests_per_second: 5.0
    burst: 10
  circuit_breaker:
    failure_threshold: 5
    recovery_timeout_seconds: 30
    half_open_max_calls: 3
  batch:
    chunk_size: 10
    max_concurrency: 3

Environment Variables

Variable	Required	Description
MISTRAL_API_KEY	Yes	Mistral API key (default model)
OPENAI_API_KEY	For GPT models	OpenAI API key
AIRTABLE_API_KEY	For Airtable	Airtable personal access token
AIRTABLE_BASE_ID	For Airtable	Airtable base ID

API Reference

load_workflow_airtable(base_id, table_name, *, ...)

Load a workflow spec from an Airtable table. Supports view, adapter (named variant), name, defaults, clients, tools, and more.

write_workflow_results(base_id, table_name, result, *, ...)

Write workflow execution results back to an Airtable table. Creates one record per step.

load_workflow_excel(path, *, ...)

Load a workflow spec from an Excel .xlsx file. Supports sheet, adapter, name, defaults, clients, tools.

write_workflow_results_excel(result, path=None, *, sheet=None, adapter=None, spec=None, run_id=None)

Write results to an Excel file. path and sheet default to values from config/adapters.yaml. Pass spec= to include passthrough columns. Pass run_id= for a custom run ID (auto-generated if omitted).

get_config()

Get the global configuration singleton.

reload_config()

Reload configuration from YAML files.

License

MIT