DEV Community

shashank srivastava
shashank srivastava

Posted on

🐾 Building a PetCare AI Assistant with Gemini and RAG

#genai #gemini #rag #llm

🚀 Overview

As part of the Gen AI Intensive Course by Google & Kaggle, I built a PetCare AI Assistant — a conversational tool that helps pet owners get structured health advice based on their pet’s symptoms.

Rather than just prompting an LLM, I implemented a full Retrieval-Augmented Generation (RAG) pipeline using Gemini, vector search, and real veterinary PDFs. The assistant outputs actionable diagnostic advice in structured JSON format, with urgency level and recommended next steps.

💡 The Problem

Most pet owners rely on scattered or unverified sources when their pet falls ill. The goal was to build an AI assistant that could understand a pet's symptoms and generate grounded, real-time advice — powered by GenAI and veterinary documents.

✅ GenAI Capabilities Demonstrated

1.Few-shot Prompting
2.Document Understanding (PDFs)
3.Structured JSON Output
4.Retrieval-Augmented Generation (RAG)
5.Evaluation of output vs expert diagnosis

📂 Dataset & Documents

I used real-world veterinary PDFs including:

  • Indian Government's Standard Treatment Guidelines (Livestock & Pets)
  • BSN Medical Veterinary Case Study Booklet
  • Veterinary Clinical Pathology reports (academic cases)

These were processed using PyMuPDF and sentence-transformers.

import fitz  # PyMuPDF

def extract_text_from_pdf(pdf_path):
    doc = fitz.open(pdf_path)
    text = ""
    for page in doc:
        text += page.get_text()
    return text
Enter fullscreen mode Exit fullscreen mode

Once text was extracted, I chunked the paragraphs and embedded them:

from sentence_transformers import SentenceTransformer

# Chunking
chunks = [para for para in text.split(". ") if len(para) > 50]

# Embedding
embed_model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = [embed_model.encode(c, convert_to_numpy=True) for c in chunks]
Enter fullscreen mode Exit fullscreen mode

🧠 Gemini + RAG Workflow

Here’s how it works:

  1. 📝 User submits a symptom query
  2. 📚 Most relevant document chunks are retrieved using embeddings
  3. 🧠 Gemini is prompted with few-shot examples + document context
  4. 🧾 Gemini generates a structured diagnosis like: 
{
  "problem_category": "health",
  "preliminary_diagnosis": "Gastrointestinal upset likely due to spoiled food",
  "suggested_action": "Provide hydration, switch to bland diet, consult vet if persists",
  "recommended_services": ["Vet Consultation", "Diet Review"],
  "urgency": "Medium"
}
Enter fullscreen mode Exit fullscreen mode

User query → converted to vector → top chunks retrieved by cosine similarity:

from sklearn.metrics.pairwise import cosine_similarity

query_vec = embed_model.encode(user_query).reshape(1, -1)
sims = cosine_similarity(query_vec, embeddings)[0]
top_idxs = sims.argsort()[-3:][::-1]
retrieved_context = "\n".join([chunks[i] for i in top_idxs])
Enter fullscreen mode Exit fullscreen mode

🧠 Output Evaluation

To measure how close Gemini’s suggestion was to a vet plan, I used cosine similarity:

from sklearn.metrics.pairwise import cosine_similarity

gemini_vec = embed_model.encode(gemini_output)
expert_vec = embed_model.encode(vet_plan)
score = cosine_similarity([gemini_vec], [expert_vec])[0][0]
Enter fullscreen mode Exit fullscreen mode

🔍 Challenges Faced

  • Gemini often returned JSON inside code blocks — had to clean with custom parser
  • Chunking was tricky for noisy scanned PDFs
  • Retrieval quality directly affected diagnosis quality

🛠 Future Improvements

Extend support to include diet, vaccination, and grooming

🔗 Kaggle Notebook Link

https://www.kaggle.com/code/shashank0907/petcare-ai-assistant

Top comments (0)

OSZAR »