langrila

Langri-La

Langrila is an open-source third-party python package allows you to develop type-safe multi-agent in an easy way. This package is just personal project.

Motivation

Widely used agent frameworks are all excellent, but I feel like:

• Often highly abstracted, making their behavior unclear at times and it causes some limitations as well.
• Code readability has some issues derived from many decorators and special operator when we are developing agent using that framework.
• Handling an agent's state or context is a little annoying. Typically it requires special arguments in functions and the state is updated hidden by tool. This might causes confusion on treaceability of the state. Ideally, functions implemented in regular development should be usable as-is with an agent, and state scope is explained by dependencies of agents, I think.
• The arguments, specifications, and available models of providers are frequently updated. To follow that is a little cumbersome.

To address these issues, I breathed new life into langrila, which has the following features:

• Deliberately avoiding over-abstracting the module.:
  • Wrapper classes of the provider sdk are minimally designed, avoiding unnecessary unification of arguments and eliminating processing dependent on model names. Langrila gives you high-level interface to use LLM while almost all the provider's original parameters also can be accepted as it is.
• No need for special argument to inject context into tools:
  • Each argument can be explicitly injected a value without the use of special argument.
• Support for sub-agents as tools:
  • In addition to passing tools to agents, sub-agents can be passed directly to the agent. Sub-agents are dynamically converted into tools internally and controlled by the parent agent, enabling easy construction of multi-agent systems. This feature brings to you intuitive coding and readability of the multi-agent.
• Unified message-response model independent of provider APIs:
  • Langrila defines a standardized message-response model that allows multi-agent systems to be built across different providers.
• Serializable conversation history:
  • The standardized message-response model can be serialized to JSON and easily stored not only in memory but also in formats like JSON, Pickle, Azure Cosmos DB, and AWS S3.
• Type-safe structured output:
  • Inspired by PydanticAI. Args are validated when a tool is invoked by the pydantic schema validator. When an error occured in validation, agent automatically retry generating response.
• Multi-modal I/O:
  • Whatever the provider supports like image/video/pdf/audio/uri input, image/audio generation, embed texts and so on.
    • I will be rolling out support progressively.
• Others:
  • Customizable internal prompts.
  • Usage gathering for all sub-agents.
  • All we have to do to support new provider is to implement a single class in many cases.

Breaking changes

Significant breaking changes were introduced from version v0.4.1 to reborn as an agent framework. It's more like a rebuild than just an update. Please be careful to update from the previous version.

Here is a rough guide of migration.

• {Provider}FunctionalChat to Agent + {Provider}Client (like OpenAIClient) module
• response.message.content[0].text to response.contents[0].text. Same for the other modalities.
• TextContent to TextPrompt / TextResponse. Same for the other modalities.
• run() / arun() to run() / run_async() in index db modules

Prerequisites

If necessary, set environment variables to use OpenAI API, Azure OpenAI Service, Gemini API, and Claude API; if using VertexAI or Amazon Bedrock, check each platform's user guide and authenticate in advance VertexAI and Amazon Bedrock.

Supported llm providers

• OpenAI
• Azure OpenAI
• Gemini on Google AI Studio
• Gemini on VertexAI
• Claude on Anthropic
• Claude on Amazon Bedrock
• Claude on VertexAI (not tested)
• Converse API models on Amazon Bedrock

Basic usage

01.introduction.ipynb

• Generate text (non-streaming / streaming, synchronous / asynchronous)
• System prompt
• Structured output using tool calling with validation
• Customize internal prompt

02.tool_calling_and_dependency_injection.ipynb

• How to use tool call with single agent (non-streaming / streaming)
• Structured output using tool calling
• Dependency injection
• Tool result serializer

03.multi_agent.ipynb

• Orchestration of multi-agent
• State handling in langrila

04.conversation_memory.ipynb

• In-memory
• Local pickle file
• Local JSON file
• Storing S3 bucket (thanks to @kun432)
• Storing Azure Cosmos DB (thanks to @rioriost)

05.multi_modal_io.ipynb

• Image input
• PDF input
• Video input
• Audio input
• Media file input from the cloud resource (GCS)
• Image generation
• Audio generation

06.embedding_text.ipynb

• Embedding with OpenAI API
• Embedding with Gemini API

07.basic_rag.ipynb

• Qdrant (Local / Remote)
• Chroma (Local / Remote)
• Usearch

Dependencies

must

python = ">=3.10,<3.13"
matplotlib = "^3.8.0"
plotly = "^5.17.0"
numpy = "^1.26.1"
pandas = "^2.1.1"
scipy = "^1.11.3"
scikit-learn = "^1.3.2"
pydantic = "^2.10.0"
griffe = "^1.5.1"
loguru = "^0.7.3"

as needed

Langrila has various extra installation options. See the following installation section and pyproject.toml.

Installation

See extra dependencies section in pyproject.toml for more detail installation options.

For user

pip

# For OpenAI
pip install langrila[openai]

# For Gemini
pip install langrila[gemini]

# For Claude
pip install langrila[claude]

# For multiple providers
pip install langrila[openai,gemini,claude]

# With dependencies to handle specific data. Here is an example using gemini
pip install langrila[gemini,audio,video,pdf]

# With dependencies for specific platform. Here is an example using gemini on VertexAI
pip install langrila[gemini,vertexai]

# With dependencies for specific vectorDB. Here is an example using Qdrant
pip install langrila[openai,qdrant]

poetry

# For OpenAI
poetry add langrila --extras openai

# For Gemini
poetry add langrila --extras gemini

# For Claude
poetry add langrila --extras claude

# For multiple providers
poetry add langrila --extras "openai gemini claude"

# With dependencies to handle specific data. Here is an example using gemini
poetry add langrila --extras "gemini audio video pdf"

# With dependencies for specific platform. Here is an example using gemini on VertexAI
poetry add langrila --extras "gemini vertexai"

# With dependencies for specific vectorDB. Here is an example using Qdrant
poetry add langrila --extras "openai qdrant"

For developer

clone

git clone [e-mail]:taikinman/langrila.git

pip

cd langrila

pip install -e .{extra packages}

poetry

# For OpenAI
poetry add --editable /path/to/langrila/ --extras "{extra packages}"

Brief roadmap

• Error handling more
• Preparing example notebooks
• Linting and refactor
• Supporting Huggingface
• Adding test code
• Aim integration

Contributing

This project is just personal, so I don't prepare the contribution guide yet, but feel free to open issues and PRs if you have any suggestions and feedbacks.

Acknowledgement

• PydanticAI
  • Awesome project to build type-safe and robust agent in a simple way. When updating langrila, I took significant inspiration from PydanticAI.