Building a Pet Hospital AI Assistant with RAG and LLMs

01 Why Use RAG

Large language models (LLMs) are trained on massive public data, but they lack access to private, time‑sensitive, or domain‑specific knowledge, which can cause hallucinations. RAG (Retrieval‑Augmented Generation) solves this by attaching an external knowledge base, allowing the model to retrieve up‑to‑date, factual information before generating answers.

Understanding RAG

RAG combines retrieval (searching a knowledge base) with generation (LLM response). It is ideal for scenarios that require external, dynamic, or proprietary data, high factual accuracy, and the ability to update knowledge without retraining the model.

External, domain‑specific or frequently updated knowledge.

High demands on answer accuracy, traceability, and factuality.

Processing of specialized or private data.

Extending knowledge without frequent model retraining.

Key Components of a RAG System

The pipeline typically includes:

Raw data parsing (convert PDFs, images, etc. to text).

Document chunking into semantically independent pieces.

Embedding (vectorization) of each chunk.

Storing vectors and metadata in a vector database.

02 Complete Build Process

2.1 Knowledge Base

Collect domain‑specific documents such as veterinary reference books, pet medical records, or imaging data. In the demo, a text file with simulated pet records is used.

2.2 Enable Alibaba Cloud AI Search Platform

Open the AI Search service, create an API key, and note the endpoint. The platform provides ready‑made services such as document parsing, vectorization, search, and LLM inference.

2.3 Create Elasticsearch Instance

Deploy an Elasticsearch instance on Alibaba Cloud to store vectors and perform similarity search. Configure VPC, network, instance specs, and whitelist your IP.

2.4 Orchestrate RAG Components

Use the platform’s template to connect document processing, vector storage, retrieval, ranking, and LLM generation. Choose appropriate models (e.g., DeepSeek‑R1, Qwen‑Turbo) and services for each stage.

2.5 Effect Testing

Run the provided Python scripts. First, execute the offline document‑processing script to parse, chunk, embed, and write vectors to Elasticsearch.

from alibabacloud_tea_openapi.models import Config
from alibabacloud_searchplat20240529.client import Client
from alibabacloud_searchplat20240529.models import GetDocumentSplitRequest

if __name__ == '__main__':
    config = Config(
        bearer_token='OS-xxx',
        endpoint='xxx.platform-cn-shanghai.opensearch.aliyuncs.com',
        protocol='http'
    )
    client = Client(config=config)
    request = GetDocumentSplitRequest().from_map({
        "document": {"content": "这是一个测试! This is a test.", "content_type": "text"},
        "strategy": {"max_chunk_size": 300, "need_sentence": False}
    })
    response = client.get_document_split("default", "ops-document-split-001", request)
    for chunk in response.body.result.chunks:
        print(chunk.content)

Replace bearer_token and endpoint with your own credentials.

After successful indexing, run the online query script. The system will:

Vectorize the user query.

Retrieve relevant document chunks via similarity search.

Re‑rank the chunks (top‑k).

Feed the query and selected context to the LLM to generate a precise answer.

Example output shows accurate retrieval of a pet’s vaccination record and personalized health advice.

Conclusion

The RAG‑enhanced pet‑hospital AI assistant demonstrates how combining LLMs with a private knowledge base can deliver highly accurate, domain‑specific answers while keeping costs low. The approach is applicable to any industry where data privacy and up‑to‑date information are critical.