🍃 Eatsy RAG Knowledge Assistant

A Retrieval-Augmented Generation (RAG) knowledge assistant for customer support in a vegan/vegetarian food delivery business.

---

🔑 Highlights

• Knowledge Base: Mock dataset with FAQs, refund policies, delivery info, and support transcripts.
• Retriever: BM25 and Gemini embedding-based retrieval (switchable).
• LLM: Google Gemini integration.
• Pipeline: Unified rag_pipeline() function for query → retrieve → context → generate.
• Portfolio Value: Demonstrates applied AI engineering (retrieval, LLM integration, RAG design).

---

⚙️ Tech Stack

• Python (pandas, dotenv)
• LlamaIndex (BM25Retriever, VectorStoreIndex, GeminiEmbedding)
• LLMs (Gemini for text generation + embeddings)
• Custom Knowledge Base

---

🚀 Quickstart

git clone https://github.com/htree255/rag-knowledge-assistant.git
cd rag-knowledge-assistant
pip install -r requirements.txt

Add a .env file with your API key and defaults:

GOOGLE_API_KEY=your_google_key_here

# Default LLM model
MODEL_NAME=gemini-2.5-flash

# Default Embedding model
EMBED_MODEL_NAME=gemini-embedding-001

# Retrieval settings
TOP_K=3

Run the pipeline:

python app.py

Example Run:

BM25-based retrieval
---------------------------------
[User] How can I get a discount?
[Eatsy Assistant] You can get discounts by referring a friend.

[User] I want to cancel my order and get a refund.
[Eatsy Assistant] You can cancel your order within 5 minutes of placing it.

[User] I want to customize my order as I have a dietary restriction.
[Eatsy Assistant] Yes, you can absolutely customize your order due to dietary restrictions!


Embedding-based retrieval
---------------------------------
[User] How can I get a discount?
[Eatsy Assistant] You can get a discount by referring a friend.

[User] I want to cancel my order and get a refund.
[Eatsy Assistant] You can cancel an order within 5 minutes of placing it. If you cancel within this timeframe, you will be eligible for a refund.

[User] I want to customize my order as I have a dietary restriction.
[Eatsy Assistant] Yes, you can customize your order due to dietary restrictions.

---

🔀 Switching Retrieval Methods

The pipeline supports both BM25 (keyword-based) and embedding-based retrieval using the Gemini embedding model.

# BM25 retriever
result = rag_pipeline("How can I get a discount?", method="bm25")

# Embedding retriever
result = rag_pipeline("How can I get a discount?", method="embedding")

This makes it easy to compare keyword search vs semantic search in customer support scenarios.

---

📓 Observations / Learnings

We compared two approaches for working with Gemini embeddings:

🔹 Direct Google SDK (genai.Client)

from google import genai
client = genai.Client()

res = client.models.embed_content(
    model="gemini-embedding-001",
    contents="What is the meaning of life?"
)
vector = res.embedding.values

✅ Pros

• Direct access to Gemini API (no extra layer).
• Full control over embedding calls and storage.
• Great for testing and custom pipelines.

⚠️ Cons

• You must manage document chunking, vector storage, and retrieval logic manually.
• More boilerplate code for a complete RAG system.

---

🔹 LlamaIndex GeminiEmbedding

from llama_index.embeddings.gemini import GeminiEmbedding

embed_model = GeminiEmbedding(model_name="gemini-embedding-001")
vector = embed_model.get_text_embedding("What is the meaning of life?")

✅ Pros

• Integrates seamlessly with LlamaIndex retrievers (VectorStoreIndex, as_retriever).
• Handles chunking, embedding, and retrieval automatically.
• Easy to switch between BM25, embeddings, or hybrid retrieval.
• Less boilerplate for building full RAG pipelines.

⚠️ Cons

• Adds dependency on LlamaIndex.
• Less direct control over low-level API calls.

---

💡 Takeaway

• Use Google SDK directly if you want full control and are building a pipeline from scratch.
• Use LlamaIndex GeminiEmbedding if you want faster development and easy integration with BM25 + hybrid retrieval.

---

🔮 Roadmap

• Add hybrid retriever (BM25 + embeddings together).
• Add LLM Guardrails (to prevent hallucinations & off-brand responses).
• Implement an LLM Judge for automated evaluation.
• Add evaluation metrics (e.g., retrieval precision/recall, grounding score).
• Build a UI chatbot with retrieval toggle.

---