Build a Medical RAG Solution with Alibaba PAI: Step-by-Step Guide

Retrieval‑Augmented Generation (RAG) for Medical Scenarios

RAG combines information retrieval with generative AI to provide more accurate and context‑relevant answers, which is crucial for domains like healthcare where precise information supports decision‑making.

1. Build Knowledge Base with PAI‑Designer

Prepare a dataset that meets PAI‑Designer’s format requirements (e.g., CSV containing disease descriptions) and import it using the Designer’s data‑import tools.

wget https://atp-modelzoo-sh.oss-cn-shanghai.aliyuncs.com/release/solutions/rag/data/%E5%8C%BB%E7%96%97csv.zip

2. Create Template in PAI‑LangStudio

In LangStudio, select the built‑in RAG template, give the application flow a name, and specify the OSS bucket that stores the dataset.

3. Deploy LLM and Embedding Models

From the Model Gallery, choose an instruction‑tuned large language model (e.g., Tongyi‑Qianwen 2.5‑7B‑Instruct) and a suitable embedding model (e.g., bge‑large‑zh‑v1.5), then deploy them. Record the VPC address and token for each service.

4. Create LLM Connection

In LangStudio’s Connection Management, add a new connection for the deployed LLM using the recorded VPC URL as base_url and the token as api_key.

5. Create Embedding Model Connection

Similarly, add a connection for the embedding service, using its VPC address and token.

6. Create Vector Database Connection (Milvus)

Configure a Milvus connection with the following key parameters: uri: the internal Milvus address, e.g.,

http://c-b1c5222fba****-internal.milvus.aliyuncs.com

token

:

<username>:<password>

database

:

default

7. PAI‑Designer Workflow for Knowledge Base

The workflow consists of four sequential components:

Read data from the OSS bucket.

Parse and chunk the text.

Generate embeddings for each chunk.

Store the vectors in the Milvus index.

8. Build Online Application in PAI‑LangStudio

After configuring the above connections, create a new application flow, add the RAG nodes (rewrite_question, retrieve, threshold_filter, generate_answer), and bind each node to the corresponding LLM, embedding, and Milvus connections.

Start the runtime, select a machine type, configure the VPC link, and deploy the RAG application.

Interact with the deployed service via the chat interface to ask medical questions.

9. Case Comparison

Task 1 – Disease Treatment Advice : The non‑RAG answer provides a generic list of therapies, while the RAG‑enhanced answer offers a more concise, clinically relevant recommendation.

Task 2 – Auxiliary Diagnosis : The non‑RAG response lists many possible conditions; the RAG‑augmented response narrows the focus to the most probable diagnosis and suggests targeted examinations.

Task 3 – Disease Early Warning : The non‑RAG answer gives a broad overview of pre‑eclampsia causes; the RAG‑enhanced answer highlights specific immune and placental mechanisms, reflecting deeper domain knowledge.