6. Graphs: Simple RAG, Pipeline RAG, and Universal RAG#

This guide provides a deep dive into Maeser’s Retrieval‑Augmented Generation (RAG) graphs—Simple RAG, Pipeline RAG, and Universal RAG—with guidance on when to use each graph. By the end of this guide, you’ll know when and how to choose each approach.

The following is a good rule of thumb for most use cases:

If your application only uses one vector store, use the Simple RAG approach.
If your application uses more than one vector store, use the Universal RAG approach.

This guide provides a description for each RAG graph but does not provide examples. For working implementations of each RAG graph, see the scripts in example/apps/ and Maeser Example (with Flask & User Management).

6.1. Prerequisites#

A Maeser development environment configured (see Development Setup).
At least one pre-built vector store (for Simple RAG) or multiple vector stores (for Pipeline or Universal RAG). Two example vector stores—byu and maeser—are provided in example/resources/vectorstores/. To create a vector store with your own content, see Embedding New Content.

6.2. Simple RAG#

The Simple RAG only takes in one vector store per chat branch, forcing the chatbot to stick to one topic per conversation.

6.2.1. When to Use Simple Rag#

Simple Rag is the best choice when:

Your application centers around one domain or subject.
You want minimal complexity and fast responses.

6.2.2. Simple RAG Framework#

        flowchart TB
    %% Nodes
    start_node(["\_\_start\_\_"])
    retrieve_context["Retrieve Relevant Context"]
    generate_response["Generate Response"]
    end_node(["\_\_end\_\_"])

    %% Main Flow
    start_node --> retrieve_context
    retrieve_context --> generate_response
    generate_response --> end_node

Retrieve Relevant Context: Scans the vector store for passages related to the user’s question and retrieves the most relevant document chunks.
Generate Response: Invokes the LLM with the conversation history, prompt instructions, and retrieved context as input, yielding a focused response.

6.2.3. Limitations of Simple RAG#

Only One Vector Store: All content used by the chatbot must be embedded into a single vector store. This will require you to compile your dataset of resources into one vector store and rebuild this vector store any time you make changes to your dataset.

6.3. Pipeline RAG#

The Pipeline RAG takes in multiple vector stores per chat branch, allowing the chatbot to dynamically choose the most relevant vector store when answering a user’s question.

Note: In almost all cases, Universal RAG is a better option compared to Pipeline RAG.

6.3.1. When to use Pipeline RAG#

Pipeline RAG is the best choice when:

Your application spans multiple knowledge bases—such as data from homework, labs, and textbooks.
Your chatbot needs to dynamically switch between knowledge bases depending on the question it is asked.

6.3.2. Pipeline RAG Framework#

        flowchart TB
    %% Nodes
    start_node(["\_\_start\_\_"])
    determine_topic["Determine Most Relevant Topic"]
    retrieve_context["Retrieve Relevant Context"]
    generate_response["Generate Response"]
    end_node(["\_\_end\_\_"])

    %% Main Flow
    start_node --> determine_topic
    determine_topic --> retrieve_context
    retrieve_context --> generate_response
    generate_response --> end_node

Determine Most Relevant Topic: Classifies the student’s question (e.g., “Is this a lab or homework question?”) to choose which vector store to query.
Retrieve Relevant Context: Scans the chosen vector store for passages related to the user’s question and retrieves the most relevant document chunks.
Generate Response: Invokes the LLM with the conversation history, prompt instructions, and retrieved context as input, yielding a focused response.

6.3.3. Limitations of Pipeline RAG#

More LLM Calls: Invokes the LLM to identify most relevant vector store before retrieving context, resulting in a slightly higher cost and response time per message (compared to Simple RAG).
One Vector Store Per Message: If the user asks a question relating to multiple vector stores, the chatbot is limited to only using one of the vector stores in its retrieval step. (Ex: If the user asks a question related to both the homework and the textbook, the chatbot can retrieve context from either the homework vector store or textbook vector store, but not both.)

6.4. Universal RAG#

Like the Pipeline RAG, the Universal RAG takes in multiple vector stores per chat branch, but unlike the Pipeline RAG, it can retrieve from multiple vector stores simultaneously, allowing the chatbot to use as many vector stores as needed to answer a user’s question.

6.4.1. When to use Universal RAG#

Universal RAG is the best choice when:

Your application spans multiple knowledge bases—such as data from homework, labs, and textbooks.
Your chatbot needs to dynamically choose which knowledge bases to pull from depending on the question it is asked.

6.4.2. Universal RAG Workflow#

        flowchart TB
    %% Nodes
    start_node(["\_\_start\_\_"])
    determine_topics["Determine Relevant Topics"]
    summarize_chat["Summarize Chat History"]
    retrieve_context["Retrieve Relevant Context"]
    generate_response["Generate Response"]
    end_node(["\_\_end\_\_"])

    %% Main Flow
    start_node --> determine_topics
    determine_topics --> summarize_chat
    summarize_chat --> |"One or More Relevant Topics"| retrieve_context
    summarize_chat --> |"No Relevant Topics"| generate_response
    retrieve_context --> generate_response
    generate_response --> end_node

Determine Relevant Topics: Classifies the student’s question and creates a list of the most relevant vector stores to query.
Summarize Chat History: Summarizes the recent chat history to provide more relevant input during the Generate Response step.
Retrieve Relevant Context: Scans each vector store in the list provided for passages related to the user’s question and retrieves the most relevant document chunks.
Generate Response: Invokes the LLM with the summarized chat history, prompt instructions, and retrieved context as input, yielding a focused response.

6.4.3. Limitations of Universal RAG#

More LLM Calls: Invokes the LLM to identify most relevant vector stores before retrieving context and summarizes chat history before generating a response, resulting in a slightly higher cost and response time per message (compared to Simple RAG).

6.5. RAG Graph Comparison Table#

Feature	Simple RAG	Pipeline RAG	Universal RAG
Vector Store Capacity	Single	Multiple	Multiple
Context Synthesis	One context	One context (chosen by relevance)	Multiple contexts (one per relevant topic)
Retrieval Steps	1	1	0 or more (1 per relevant topic)
LLM Calls per Response	2	3	3 plus number of relevant topics/contexts

6.6. Next Steps#

Review the scripts in example/apps/ and Maeser Example (with Flask & User Management) for implementations of each RAG graph.
Explore Custom Graphs for tool integration (e.g., calculators) in Custom Graphs: Advanced RAG Workflows.
Explore the source documentation for the maeser.graphs subpackage.

Graphs: Simple RAG, Pipeline RAG, and Universal RAG

Contents