Build an Education‑Focused Retrieval‑Augmented Generation (RAG) Solution with Alibaba PAI

Introduction

Retrieval‑Augmented Generation (RAG) combines information retrieval with generative AI to deliver more accurate and context‑relevant answers, which is especially valuable in education where precise information is required.

Prerequisites

Activate PAI pay‑as‑you‑go and create a default workspace (see the PAI onboarding guide).

Create an OSS bucket for storing training data.

Enable a Milvus vector database instance (see the quick‑start guide).

1. Build a Knowledge Base with PAI‑Designer

Prepare your dataset according to PAI‑Designer’s format requirements (CSV is supported). The example uses a CSV file containing biology course knowledge points from Wikipedia.

Download the sample dataset:

wget https://atp-modelzoo-sh.oss-cn-shanghai.aliyuncs.com/release/solutions/rag/data/%E6%95%99%E8%82%B2csv.zip

Upload the extracted CSV to your own OSS bucket for the next steps.

2. Deploy LLM and Embedding Models

Navigate to Model Gallery in the PAI console, select a large language model (e.g., 通义千问2.5‑7B‑Instruct ) and an embedding model (e.g., bge‑large‑zh‑v1.5 ), and deploy them. Ensure you choose an instruction‑tuned model (name contains “Chat” or “Instruct”).

3. Create LLM Connection

In LangStudio , go to Connection Management → New Connection . Fill in the VPC address and token obtained from the deployed LLM service.

4. Create Embedding Model Connection

Similarly, create a connection for the embedding model using its VPC address and token.

5. Create Vector Database Connection

Configure a Milvus connection:

uri : http://<Milvus‑internal‑address> (e.g., http://c-b1c5222fba****-internal.milvus.aliyuncs.com)

token : <username>:<password> database :

default

Knowledge‑Base Workflow in PAI‑Designer

Read data from the OSS bucket.

Parse and chunk the text.

Generate vectors for each chunk.

Store vectors in the Milvus vector database.

These steps can be assembled visually in PAI‑Designer; the workflow diagram is shown below.

Template Building in PAI‑LangStudio

Use the built‑in RAG template to create an application flow:

Enter LangStudio, select your workspace, and click New Application Flow .

Choose the RAG template, name the flow, and specify the OSS path for assets.

Application Flow Nodes

rewrite_question : Rewrites the user query for better retrieval.

retrieve : Calls the Milvus vector store to fetch relevant documents.

threshold_filter : Filters retrieved documents based on similarity score.

generate_answer : Generates the final answer using the LLM connection.

Case Comparison: RAG vs. Non‑RAG

Task 1 – Scientific Knowledge Q&A

Question: Summarize how scientists discovered meiosis.

Without RAG (baseline LLM): Provides a lengthy answer with some factual inaccuracies and unnecessary details.

With RAG: Returns a concise, accurate summary that correctly attributes discoveries to Oscar Hertwig, Edouard Van Beneden, Theodor Boveri, Thomas Hunt Morgan, and later molecular studies.

Task 2 – Intelligent Grading and Feedback

Question: Identify errors in a statement about the origin of photosynthesis.

Without RAG: Gives a generic critique with limited detail.

With RAG: Highlights specific inaccuracies: chlorophyll absorbs blue/red light, diversity of pigments, debated timing of origin (~3.5 Ga), the role of cyanobacteria, and the correct primary products of photosynthesis.

Running the Application

After configuring the flow, start the runtime, select a machine type, and deploy the RAG service. Use the built‑in chat interface to ask questions and receive RAG‑enhanced answers.