advanced-rag

The repository structure should reflect the architecture of the project, which represent the nodes and edges implemented.

%%{init: {'flowchart': {'curve': 'linear'}}}%%
graph TD;
	__start__([<p>__start__</p>]):::first
	retrieve(retrieve)
	grade_documents(grade_documents)
	generate(generate)
	websearch(websearch)
	__end__([<p>__end__</p>]):::last
	__start__ --> retrieve;
	generate --> __end__;
	retrieve --> grade_documents;
	websearch --> generate;
	grade_documents -.-> websearch;
	grade_documents -.-> generate;
	classDef default fill:#f2f0ff,line-height:1.2
	classDef first fill-opacity:0
	classDef last fill:#bfb6fc

Loading

• graph.py - Connects all nodes and edges.
• state.py - Graph state objects that will be modified during the Graph execution.
• consts.py - Constants used in the implementation.
• nodes/ - Contains the implementation of all nodes that will run.
• chains/ - Each file will be a different chain that will correspond to a node (each node will run a LangChain chain).
• tests/ - Contains the tests for the chains.
• ingestion.py - Hold the logic that will download information to be indexed in the vector store

Self RAG

Reflect on the answer and compare it to the documents to check if the model hallucinated. Check if the result is grounded in the documents.

%%{init: {'flowchart': {'curve': 'linear'}}}%%
graph TD;
	__start__([<p>__start__</p>]):::first
	retrieve(retrieve)
	grade_documents(grade_documents)
	generate(generate)
	websearch(websearch)
	__end__([<p>__end__</p>]):::last
	__start__ --> retrieve;
	generate --> __end__;
	retrieve --> grade_documents;
	websearch --> generate;
	grade_documents -.-> websearch;
	grade_documents -.-> generate;
	generate -. &nbsp;useful&nbsp; .-> __end__;
	generate -. &nbsp;not useful&nbsp; .-> websearch;
	generate -. &nbsp;not supported&nbsp; .-> generate;
	classDef default fill:#f2f0ff,line-height:1.2
	classDef first fill-opacity:0
	classDef last fill:#bfb6fc

Loading

Adaptive RAG

Fancy word for simply using a question router to route our prompt to different RAG flows.

%%{init: {'flowchart': {'curve': 'linear'}}}%%
graph TD;
	__start__([<p>__start__</p>]):::first
	retrieve(retrieve)
	grade_documents(grade_documents)
	generate(generate)
	websearch(websearch)
	__end__([<p>__end__</p>]):::last
	retrieve --> grade_documents;
	websearch --> generate;
	__start__ -.-> websearch;
	__start__ -.-> retrieve;
	grade_documents -.-> websearch;
	grade_documents -.-> generate;
	generate -. &nbsp;useful&nbsp; .-> __end__;
	generate -. &nbsp;not useful&nbsp; .-> websearch;
	generate -. &nbsp;not supported&nbsp; .-> generate;
	classDef default fill:#f2f0ff,line-height:1.2
	classDef first fill-opacity:0
	classDef last fill:#bfb6fc

Loading

Running the project

Uncomment the vectorstore code at the ingestion file and run the project. A new .chroma folder will be created at the root of the application.

Once the vector data is generated, the vectorstore command can be commented out in order to avoid the same data ingestion everytime the project runs.

Utilities

Installing Dependencies:

poetry add beautifulsoup4 langchain langgraph langchainhub langchain-community tavily-python langchain-chroma langchain_openai python-dotenv black isort pytest grandalf

Running tests:

pytest . -s -v

Environment variables - .env file:

OPENAI_API_KEY=<>
TAVILY_API_KEY=<>
LANGCHAIN_API_KEY=<>
LANGCHAIN_TRACING_V2=true
LANGCHAIN_PROJECT=CRAG
PYTHONPATH=</Users/../advanced-rag>

Reference

• Advanced RAG control flow with Mistral and LangChain
• LangChain's Cookbook
• langgraph-course repository from Eden Marco
• LangGraph- Develop LLM powered AI agents with LangGraph
• Self-RAG: Learning to retrieve, generate, and critique through self-reflection
• Adaptive RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity