orac 0.1.0

A lightweight, YAML-driven framework for OpenAI-compatible LLM APIs.

Orac

Orac is a lightweight, YAML-driven framework for working with OpenAI-compatible LLM APIs. It provides clean abstractions, command-line integration, structured parameter handling, and support for both local and remote file attachments.

---

Features

• Prompt-as-config: Define entire LLM tasks in YAML, including prompt text, parameters, default values, model settings, and file attachments.
• Hierarchical configuration: Three-layer config system (base → prompt → runtime) with deep merging for flexible overrides.
• Templated inputs: Use ${variable} placeholders in prompt and system prompt fields.
• File support: Attach local or remote files (e.g., images, documents) via files: or file_urls: in YAML or CLI flags.
• Command-line and Python API: Use either the CLI tool or the LLMWrapper class in code.
• Runtime configuration overrides: Override model settings, API keys, generation options, and safety filters from the CLI or programmatically.
• Structured output support: Request application/json responses or validate against a JSON Schema.
• Parameter validation: Automatically convert and validate inputs by type.
• Logging: Logs all operations to file and provides optional verbose console output.

---

Installation

Option 1: Using requirements.txt

pip install -r requirements.txt

Option 2: Manual installation

pip install google-generativeai openai PyYAML python-dotenv loguru

---

Configuration

Environment Variables

Orac supports configuration through environment variables. You can either set them directly or use a .env file:

1. Copy the example environment file:

   cp .env.example .env
   

2. Edit .env with your settings:

   # API Keys
   GOOGLE_API_KEY=your_google_api_key_here
   OPENAI_API_KEY=your_openai_api_key_here
   
   # Configuration overrides (optional)
   ORAC_DEFAULT_MODEL_NAME=gemini-2.0-flash
   ORAC_LOG_FILE=./llm.log
   

3. Or set environment variables directly:

   export ORAC_LLM_PROVIDER="google"
   export GOOGLE_API_KEY="your_api_key_here"
   export ORAC_DEFAULT_MODEL_NAME="gemini-2.0-flash"
   

Choosing an LLM Provider

Orac requires explicit provider selection. You must specify which LLM provider to use either via environment variable or CLI flag:

Provider	ORAC_LLM_PROVIDER	API Key Environment Variable	Default Base URL
Google Gemini	google	GOOGLE_API_KEY	https://generativelanguage.googleapis.com/v1beta/openai/
OpenAI	openai	OPENAI_API_KEY	https://api.openai.com/v1/
Anthropic	anthropic	ANTHROPIC_API_KEY	https://api.anthropic.com/v1/
Azure OpenAI	azure	AZURE_OPENAI_KEY	${AZURE_OPENAI_BASE} (user-set)
OpenRouter	openrouter	OPENROUTER_API_KEY	https://openrouter.ai/api/v1/
Custom	custom	user picks	user sets via --base-url

Examples:

# Using Google Gemini
export ORAC_LLM_PROVIDER=google
export GOOGLE_API_KEY=your_google_api_key
python -m orac capital --country France

# Using OpenAI
export ORAC_LLM_PROVIDER=openai
export OPENAI_API_KEY=your_openai_api_key
python -m orac capital --country Spain

# Using OpenRouter (access to multiple models)
export ORAC_LLM_PROVIDER=openrouter
export OPENROUTER_API_KEY=your_openrouter_api_key
python -m orac capital --country Japan

# Using CLI flags instead of environment variables
python -m orac capital --provider google --api-key your_api_key --country Italy

# Using a custom endpoint
python -m orac capital --provider custom --base-url https://my-custom-api.com/v1/ --api-key your_key --country Germany

Configurable Environment Variables

All default settings can be overridden with environment variables using the ORAC_ prefix:

• ORAC_LLM_PROVIDER - Required: LLM provider selection (google|openai|anthropic|azure|openrouter|custom)
• ORAC_DEFAULT_MODEL_NAME - Default LLM model
• ORAC_DEFAULT_PROMPTS_DIR - Directory for prompt files
• ORAC_DEFAULT_CONFIG_FILE - Path to config YAML
• ORAC_DOWNLOAD_DIR - Temp directory for file downloads
• ORAC_LOG_FILE - Log file location

Configuration Hierarchy

Orac uses a layered configuration system, allowing for flexible and powerful control over your prompts. Settings are resolved with the following order of precedence (where higher numbers override lower ones):

1. Base Configuration (orac/config.yaml): The default settings for the entire project. This file is included with the orac package and provides sensible defaults for model_name, generation_config, and safety_settings. You can edit it directly in your site-packages or provide your own via a custom script.

2. Prompt Configuration (prompts/your_prompt.yaml): Any setting defined in a specific prompt's YAML file will override the base configuration. This is the primary way to customize a single task. For example, you can set a lower temperature for a factual prompt or a different model_name for a complex one.

3. Runtime Overrides (CLI / Python API): Settings provided directly at runtime, such as using the --model-name flag in the CLI or passing the generation_config dictionary to the LLMWrapper constructor, will always take the highest precedence, overriding all other configurations.

Example Override

If orac/config.yaml has:

# orac/config.yaml
generation_config:
  temperature: 0.7

And your prompt has:

# prompts/recipe.yaml
prompt: "Give me a recipe for ${dish}"
generation_config:
  temperature: 0.2  # Override for more deterministic recipes

Running orac recipe will use a temperature of 0.2.

Running orac recipe --generation-config '{"temperature": 0.9}' will use a temperature of 0.9.

---

Example Usage

1. Create a YAML prompt

Save the following to prompts/capital.yaml:

prompt: "What is the capital of ${country}?"
parameters:
  - name: country
    description: Country name
    default: France

2. Run from Python

from orac import LLMWrapper

llm = LLMWrapper("capital")
print(llm.completion())  # Defaults to France
print(llm.completion(country="Japan"))

3. Run from CLI

orac capital
orac capital --country Japan
orac capital --verbose
orac capital --info

4. Advanced examples

# Override model and config
orac capital --country "Canada" \
  --model-name "gemini-2.0-flash" \
  --generation-config '{"temperature": 0.4}'

# Structured JSON response
orac recipe --json-output

# Schema validation
orac capital --country "Germany" \
  --response-schema schemas/capital.schema.json

# Attach local and remote files
orac paper2audio \
  --file reports/report.pdf \
  --file-url https://example.com/image.jpg

---

YAML Prompt Reference

Basic YAML

prompt: "Translate the following text: ${text}"
parameters:
  - name: text
    type: string
    required: true

Additional Options

model_name: gemini-2.0-flash
api_key: ${OPENAI_API_KEY}

generation_config:
  temperature: 0.5
  max_tokens: 300

safety_settings:
  - category: HARM_CATEGORY_HARASSMENT
    threshold: BLOCK_NONE

response_mime_type: application/json
response_schema:
  type: object
  properties:
    translation: { type: string }

files:
  - data/*.pdf
file_urls:
  - https://example.com/image.jpg

require_file: true

Supported Parameter Types

• string
• int
• float
• bool
• list (comma-separated values)

---

CLI Options

orac <prompt_name> [--parameter-name VALUE ...] [options]

Global Flags

• --info: Show parameter metadata
• --verbose, -v: Enable verbose logging
• --prompts-dir DIR: Use custom prompt directory
• --model-name MODEL
• --api-key KEY
• --generation-config JSON
• --safety-settings JSON
• --file FILE
• --file-url URL
• --json-output
• --response-schema FILE
• --output FILE, -o

---

Logging

Orac provides comprehensive logging with two output modes:

Default Mode (Quiet)

• Only shows LLM responses and critical errors
• All detailed logging goes to file only
• Perfect for clean integration and scripting

Verbose Mode

• Shows detailed operation logs on console
• Includes timestamps, function names, and colorized output
• Enable with --verbose or -v flag

Log Configuration

• File logging: All activity logged to llm.log (configurable via ORAC_LOG_FILE)
• Rotation: 10 MB max file size, 7 days retention
• Levels: DEBUG level in files, INFO+ in console (verbose mode)

Usage Examples

# Quiet mode (default) - only shows LLM response
orac capital --country France

# Verbose mode - shows detailed logging
orac capital --country Spain --verbose

# Check recent logs
tail -f llm.log

To configure logging programmatically:

from orac.logger import configure_console_logging
configure_console_logging(verbose=True)

---

Development & Testing

To run the test suite:

python test.py