ai_api_unified 0.1.2

Unified access layer for completion and embedding services

ai_api_unified · Unified Foundation-Model Client Library

Version: 0.1.0  |  License: MIT

ai_api_unified provides a single, typed interface for calling both completion-style LLMs and text-embedding models across vendors (OpenAI, Amazon Bedrock/Titan, …).

Prerequisites

• Python 3.12.1 (only)
  We strongly recommend using pyenv to install and pin exactly 3.12.1, so that compiled wheels (e.g. tiktoken) are available and no Rust toolchain is required.

Structure

classDiagram

%% Abstract Base Classes
class AIBase {
  <<abstract>>
  +list_model_names: List[str]
  +count_tokens(text: str): int
}

class AIBaseEmbeddings {
  <<abstract>>
  +list_model_names: List[str]
  +generate_embeddings(text: str): Dict[str, Any]
  +generate_embeddings_batch(texts: List[str]): List[Dict[str, Any]]
}

class AIBaseCompletions {
  <<abstract>>
  +list_model_names: List[str]
  +max_context_tokens: int
  +price_per_1k_tokens: float
  +strict_schema_prompt(prompt: str, response_model: Type[AIStructuredPrompt], max_response_tokens: int): AIStructuredPrompt
  +send_prompt(prompt: str): str
}

class AIStructuredPrompt {
  <<abstract>>
  +prompt: str
  +get_prompt(): Optional[str]
  +send_structured_prompt(ai_client: AIBaseCompletions, response_model: Type[AIStructuredPrompt]): Optional[AIStructuredPrompt]
}

%% Concrete Embedding Classes
class AiOpenAIEmbeddings {
  +generate_embeddings(text: str): Dict[str, Any]
  +generate_embeddings_batch(texts: List[str]): List[Dict[str, Any]]
}

class AiTitanEmbeddings {
  +generate_embeddings(text: str): Dict[str, Any]
  +generate_embeddings_batch(texts: List[str]): List[Dict[str, Any]]
}

%% Concrete Completion Classes
class AiOpenAICompletions {
  +strict_schema_prompt(prompt: str, response_model: Type[AIStructuredPrompt], max_response_tokens: int): AIStructuredPrompt
  +send_prompt(prompt: str): str
}

class AiBedrockCompletions {
  +strict_schema_prompt(prompt: str, response_model: Type[AIStructuredPrompt], max_response_tokens: int): AIStructuredPrompt
  +send_prompt(prompt: str): str
}

%% Utility Classes
class EnvSettings {
  +get_setting(setting: str, default: Any): Any
  +is_setting_on(setting: str): bool
  +override_setting(setting: str, value: Any): None
}

%% Factory Class
class AIFactory {
  +get_ai_client(client_type: str): AIBase
  +get_ai_completions_client(client_type: str, model_name: Optional[str]): AIBaseCompletions
  +get_ai_embedding_client(client_type: str): AIBaseEmbeddings
}

%% Relationships
AIBase <|-- AIBaseEmbeddings
AIBase <|-- AIBaseCompletions

AIBaseEmbeddings <|-- AiOpenAIEmbeddings
AIBaseEmbeddings <|-- AiTitanEmbeddings

AIBaseCompletions <|-- AiOpenAICompletions
AIBaseCompletions <|-- AiBedrockCompletions

AIStructuredPrompt <|-- ExampleStructuredPrompt : inherits

AIFactory --> AIBase : creates
AIFactory --> AIBaseCompletions : creates
AIFactory --> AIBaseEmbeddings : creates

AiOpenAIEmbeddings --> EnvSettings : uses
AiTitanEmbeddings --> EnvSettings : uses
AiOpenAICompletions --> EnvSettings : uses
AiBedrockCompletions --> EnvSettings : uses

Installation

# from your internal Artifactory PyPI
pip install --index-url https://<org>.jfrog.io/artifactory/api/pypi/pypi-local/simple ai_api_unified

Quick start

Configuration via Environment Variables

This library is fully configurable through environment variables—in exactly the way shown in the provided .env_template. You can swap providers, engines, and models without changing code. Below are the key variables:

EMBEDDING_ENGINE

Controls which embedding backend to use.

• openai → uses OpenAI’s embedding API
• titan → uses Amazon Titan embeddings

Default: openai

COMPLETIONS_ENGINE

Controls which family of completion (chat/LLM) models to use.

• openai → uses OpenAI’s completion API
• nova → uses Amazon Nova (Bedrock) LLMs
• llama → uses Meta Llama family via Bedrock
• anthropic → uses Anthropic Claude family via Bedrock
• mistral → uses Mistral family via Bedrock
• cohere → uses Cohere Command family via Bedrock
• ai21 → uses AI21 Jamba family via Bedrock
• rerank → uses Amazon Rerank via Bedrock

Default: openai

Submit a PR to add new ones.

OPENAI_API_KEY

Your OpenAI API key. Required when EMBEDDING_ENGINE or COMPLETIONS_ENGINE is set to openai.

EMBEDDING_MODEL_NAME

The exact model identifier for embeddings.

• OpenAI examples:

  • text-embedding-3-small – cost-effective small embedding
  • text-embedding-3-large – higher-capacity large embedding
  • text-embedding-ada-002 – versatile general-purpose embedding
  • text-search-ada-doc-001 – optimized for document search
  • text-search-davinci-doc-001 – high-accuracy document search
  • code-search-ada-text-001 – code-aware embedding for search
  • code-search-babbage-text-001 – larger-capacity code search
  • text-similarity-ada-001 – basic text similarity
  • text-similarity-babbage-001 – mid-range text similarity
  • text-similarity-curie-001 – higher-quality text similarity

• Amazon Bedrock examples:

  • amazon.titan-embed-text-v2:0 – Titan Text Embed V2
  • amazon.titan-embed-text-v1:0 – Titan Text Embed V1

Default: text-embedding-3-small

COMPLETIONS_MODEL_NAME

The exact model identifier for completions. Changing this in config will allow a no-code update to your program. Alternatively, you can set the model on class initialization in code.

• OpenAI example:

  • gpt-4o-mini – optimized for cost-sensitive, low-latency use
  • gpt-4o – general-purpose GPT-4o
  • gpt-4o-16k – GPT-4o with 16 000-token context
  • gpt-4 – standard GPT-4
  • gpt-4-32k – GPT-4 with 32 000-token context
  • gpt-3.5-turbo – flagship GPT-3.5 model
  • gpt-3.5-turbo-16k – GPT-3.5 with 16 000-token context
  • text-davinci-003 – high-quality text generation
  • text-curie-001 – balanced speed and capability
  • code-davinci-002 – code-optimized completions

• Amazon Bedrock examples:

  • amazon.nova-pro-v1:0 – Nova Pro
  • amazon.nova-lite-v1:0 – Nova Lite
  • amazon.nova-micro-v1:0 – Nova Micro
  • amazon.nova-canvas-v1:0 – Nova Canvas
  • amazon.titan-text-premier-v1:0 – Titan Text Premier
  • anthropic.claude-opus-4-20250514-v1:0 – Claude Opus 4
  • anthropic.claude-sonnet-4-20250514-v1:0 – Claude Sonnet 4
  • meta.llama2-70b-chat-hf:2 – Llama 2 Chat 70B
  • meta.llama3-70b-instruct-v1:0 – Llama 3 Instruct 70B
  • mistral.mistral-large-2407-v1:0 – Mistral Large
  • cohere.command-r-plus-v1:0 – Cohere Command R+
  • ai21.jamba-1-5-large-v1:0 – AI21 Jamba 1.5 Large

Default: gpt-4o-mini

EMBEDDING_DIMENSIONS

Dimensionality of the embedding vectors.

• 1536 for OpenAI embeddings
• 1024 for Titan embeddings

Default: 1536

AWS_REGION

AWS region for Bedrock/Titan when using Amazon services. Note: to use Bedrock, you must get your access environment set up independently, since there is no API key approach to using Bedrock.

Default: us-east-1

Example Code

"""
Creates a foundation model API client with a single call.
Can swap models with a config or param change.

Create a structured output subclass with your required structure,
and then use it with a single call.
"""
from ai_api_unified import AIFactory, AIStructuredPrompt

# Simple Completions
client = AIFactory.get_ai_completions_client()           # auto-selects engine via .env
response = client.send_prompt("Say hello in German")
print(response)  # → "Hallo!"

# Structured Prompts
class YourStructuredPrompt(AIStructuredPrompt):
    message_input_field: str  # this is an input field, not a result

    message_output_field: Optional[str] = None # This is a parsed output field


    @staticmethod
    def get_prompt(
        message_input: str,
    ) -> str:
        prompt = textwrap.dedent(
            f"""
            Reply with than uppercase version of the message_input in the test_output field.
            message_input: '{message_input_field}'
            """
        ).strip()
        return prompt

    @classmethod
    def model_json_schema(cls) -> Dict[str, Any]:
        """
        JSON schema for the LLM’s *output* only.
        """
        # start with a fresh copy of the base schema (deep-copied there)
        schema: Dict[str, Any] = deepcopy(super().model_json_schema())
        schema["properties"]["message_output_field"] = {"type": "string"}
        # make test_output required for the LLM response
        schema.setdefault("required", [])
        schema["required"].append("message_output_field")
        return schema


  structured_prompt: YourStructuredPrompt = YourStructuredPrompt(message_input_field="hello")
  structured_prompt_result: YourStructuredPrompt = (
      structured_prompt.send_structured_prompt(
          client, ExampleStructuredPrompt
      )
  )
  print(structured_prompt_result.message_output_field) # -> "HELLO"

# Embeddings client use
embedder: AIBaseEmbeddings = AIFactory.get_ai_embedding_client()
dict_embeddings: Dict[str, Any] = embedder.generate_embeddings("vectorize me")
# dict_embeddings holds {"embedding": [], "text": "vectorize me", "dimensions": int}

---

Repository layout

src/ai_api_unified/          ← package source
└── ai_base.py           ← abstract interfaces
└── ai_factory.py        ← runtime factory
tests/                   ← pytest suite
.env_template            ← sample environment config

---

Development

# create virtualenv & install runtime + dev dependencies
poetry install --with dev
pytest -q

Roadmap

• Add simple method for a developer to cycle through models to test them
• Add more provider back-ends (Anthropic, Google).
• Provide async variants for high-throughput workloads.