How snowflake-labs/arctic_agentic_rag works
Project Type and Purpose
• This is an enterprise-focused, modular, retrieval-augmented generation (RAG) framework designed to handle ambiguous queries and deliver smarter, faster, and more reliable AI-driven workflows.
• It is built around agent-based modular design with multiple integrations for various LLM backends (including Azure OpenAI, Cortex, Ollama, and vLLM) and incorporates retrieval and ranking capabilities.
• The framework also supports YAML-based configuration and comes with prebuilt examples and demo projects.
File Structure to Extract Core Components
• The core package (“arctic_agentic_rag”) is organized into several submodules:
• Agent Module (arctic_agentic_rag/agent): Contains the agent logic driving the agentic workflow.
• LLM Module (arctic_agentic_rag/llm): Manages integrations with different LLM providers (azure_oai.py, cortex_complete.py, oai.py, ollama.py, vllm_provider.py). This is your bridge to various external AI services.
• Ranking Module (arctic_agentic_rag/ranking): Includes functionality for clustering and encoding to rank or cluster results.
• Retrieval Module (arctic_agentic_rag/retrieval): Provides the base interface plus implementations, such as cortex_search and qwen_retrieval for fetching information.
• Template Module (arctic_agentic_rag/template): Supports dynamic template creation/field templating for prompt engineering.
• Utility and Logging Modules: Provide cross-cutting support like logging, helper functions, and miscellaneous utilities.
• Additional directories like examples and projects demonstrate concrete use cases (e.g., ambiguous query resolution in Episode 1 and simple RAG demos).
• Configuration files (YAML in examples and projects) define how to set up pipelines and experiments.