language-model-common 2.0.21

Provides the underlying framework to enhance langchain and add loading configurations

language-model-common

A shared Python framework for building LLM-powered agent applications with LangChain and LangGraph. It provides reusable infrastructure for configuration management, protocol conversion, MCP tool integration, file management, and more.

Features

• Multi-source configuration loading — Read LLM model configs from local filesystem, AWS S3, or GitHub repositories with TTL-based caching and client-specific overrides
• LangGraph-to-OpenAI protocol conversion — Stream LangGraph agent output as OpenAI-compatible Server-Sent Events (SSE) for chat completion APIs
• MCP (Model Context Protocol) integration — Tool discovery with BM25 search ranking, OAuth 2.1/OIDC support, dynamic client registration, and MCP Apps UI rendering via ui:// resources
• Prompt template library — Load and manage prompt templates from organized directory structures with GitHub auto-download support
• File management abstraction — Unified interface for local and AWS S3 storage with factory-based backend selection
• Token and cost management — Token reduction for long conversations and usage metadata tracking via tiktoken
• Authentication and authorization — OAuth with PKCE, JWT token validation, OIDC provider discovery
• Image generation — Provider abstraction over OpenAI and AWS Bedrock image generators
• OCR extraction — AWS Textract integration behind a factory interface
• Dependency injection container — Pre-wired service registry using simple-container with singleton lifecycle management

Installation

pip install language-model-common

Requirements: Python >= 3.10

Quick Start

Reading Model Configurations

from languagemodelcommon.configs.config_reader import ConfigReader

config_reader = ConfigReader(
    config_paths=["./configs"],  # local path, s3:// URI, or github:// path
)

# Load base configs with optional client-specific overrides
models = await config_reader.read_model_configs_async(client_id="my-client")

Configuration sources are selected by path prefix:

• Local: ./configs or /absolute/path
• S3: s3://bucket-name/path
• GitHub: Managed via GithubConfigRepoManager with automatic background refresh

Loading Prompts

from languagemodelcommon.configs.prompt_library import PromptLibraryManager

prompt_manager = PromptLibraryManager(config_paths=["./configs"])
prompt_text = prompt_manager.get_prompt("system-prompt")

Prompts are loaded from .md or .txt files in a prompts/ subdirectory of your config paths.

Streaming LangGraph Output as OpenAI SSE

from languagemodelcommon.converters import LangGraphToOpenAIConverter

converter = LangGraphToOpenAIConverter(
    streaming_manager=streaming_manager,
    token_reducer=token_reducer,
)

response = await converter.stream_response(
    graph=my_langgraph,
    messages=messages,
    config=config,
)

File Management

from languagemodelcommon.file_managers import FileManagerFactory

factory = FileManagerFactory(aws_client_factory=aws_factory)

# Automatically selects local or S3 backend based on the folder path
manager = factory.create("s3://my-bucket")
await manager.save_file_async(file_data, "s3://my-bucket", "output.json", "application/json")

MCP Tool Discovery

from languagemodelcommon.mcp import ToolCatalog

catalog = ToolCatalog()
catalog.register_server("my-server", url="http://localhost:8080")

# BM25-ranked search for relevant tools
results = catalog.search("search patient records")

Using the Dependency Injection Container

from languagemodelcommon.container import LanguageModelCommonContainerFactory
from simple_container import SimpleContainer

container = SimpleContainer()
LanguageModelCommonContainerFactory.register_services_in_container(container)

config_reader = container.resolve(ConfigReader)

Environment Variables

Variable	Description	Default
GITHUB_CONFIG_REPO_URL	GitHub zipball API URL for config repo	—
GITHUB_TOKEN	GitHub token for authenticated requests	—
GITHUB_CACHE_FOLDER	Local cache directory for GitHub configs	/tmp/github_config_cache
CONFIG_CACHE_TIMEOUT_SECONDS	Config refresh interval (seconds)	120
GITHUB_TIMEOUT	HTTP timeout for GitHub requests (seconds)	300

Development

Prerequisites

• Docker and Docker Compose

Setup

make init

This builds the development Docker image, locks dependencies, and sets up pre-commit hooks.

Running Tests

make tests

Tests run inside Docker using pytest with async support (asyncio_mode = auto).

Other Commands

make shell       # Open a shell in the dev container
make update      # Update dependencies and rebuild
make build       # Build distribution package
make package     # Publish to PyPI
make testpackage # Publish to TestPyPI

Project Structure

languagemodelcommon/
├── auth/              # Token storage and auth managers
├── aws/               # AWS client factory
├── configs/           # Config reader, prompt library, schemas
├── container/         # DI container factory
├── converters/        # LangGraph-to-OpenAI conversion, streaming
├── exceptions/        # Custom exception types
├── file_managers/     # Local and S3 file management
├── graph/             # LangGraph utilities
├── history/           # Conversation history management
├── http/              # HTTP client factory
├── image_generation/  # Image generation providers
├── markdown/          # HTML/CSV to Markdown converters
├── mcp/               # MCP client, tool catalog, OAuth
├── models/            # LLM model definitions
├── mocks/             # Test mocks and fakes
├── ocr/               # OCR extraction (AWS Textract)
├── persistence/       # LangGraph checkpoint/store backends
├── schema/            # OpenAI-compatible schema definitions
├── state/             # LangGraph state definitions
├── structures/        # Request/response wrappers
├── tools/             # Resilient tool base class, MCP tools
└── utilities/         # Logging, caching, token reduction, security

License

Apache License 2.0