Langroid

Agents serve as core abstractions, encapsulating LLM conversation state, memory (vector-stores), and tools.

Defines tasks that wrap agents with instructions, enabling hierarchical and recursive task delegation among agents.

Promotes flexible design of specialized agents and tasks, allowing for their easy combination and reuse in complex applications.

Seamlessly integrates with various LLMs, including local, open-source, proprietary, and API-based models.

Supports multiple vector stores (e.g., Qdrant, Chroma, LanceDB) to facilitate Retrieval-Augmented Generation (RAG).

Simplifies defining and using tools/functions with Pydantic, ensuring structured output and improved error handling for LLM interactions.

Leverages Redis for caching LLM prompts and responses, enhancing performance and reducing operational costs.

Enables grounding of LLM responses by accessing external documents and provides mechanisms for source citation.

Provides comprehensive logs of multi-agent interactions and maintains message provenance for full observability.