rag

RAG (Retrieval-Augmented Generation) Examples

This directory contains examples of implementing RAG systems with Mellea.

Files

simple_rag_with_filter.py

A complete RAG pipeline with relevance filtering using generative stubs.

Key Features:

• Vector embedding with FAISS index
• Semantic search for document retrieval
• LLM-based relevance filtering
• Answer generation with grounded context
• Using @generative for filtering logic

mellea_pdf.py

RAG example specifically for PDF documents.

Concepts Demonstrated

• Vector Search: Using embeddings for semantic retrieval
• Relevance Filtering: Using LLMs to filter retrieved documents
• Grounded Generation: Generating answers based on retrieved context
• Multi-Stage Pipeline: Retrieval → Filtering → Generation
• Generative Stubs: Using @generative for RAG components

Pipeline Architecture

Query
  ↓
Embedding Model (sentence-transformers)
  ↓
Vector Search (FAISS)
  ↓
Retrieved Documents
  ↓
Relevance Filter (LLM)
  ↓
Relevant Documents
  ↓
Answer Generation (LLM with grounding_context)
  ↓
Final Answer

Basic Usage

from mellea import generative, start_session
from sentence_transformers import SentenceTransformer
from faiss import IndexFlatIP

# 1. Create embeddings and index
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = embedding_model.encode(documents)
index = IndexFlatIP(dimension)
index.add(embeddings)

# 2. Retrieve documents
query_embedding = embedding_model.encode([query])
distances, indices = index.search(query_embedding, k=5)
retrieved_docs = [documents[i] for i in indices[0]]

# 3. Filter for relevance
@generative
def is_answer_relevant_to_question(answer: str, question: str) -> bool:
    """Determine whether the answer is relevant to the question."""

m = start_session()
relevant_docs = [
    doc for doc in retrieved_docs 
    if is_answer_relevant_to_question(m, answer=doc, question=query)
]

# 4. Generate answer with grounding
answer = m.instruct(
    "Answer the question: {{query}}",
    user_variables={"query": query},
    grounding_context={f"doc{i}": doc for i, doc in enumerate(relevant_docs)}
)

Dependencies

# Install RAG dependencies
uv pip install faiss-cpu sentence-transformers

Key Techniques

• Semantic Search: Finding relevant documents by meaning, not keywords
• Two-Stage Retrieval: Broad retrieval followed by precise filtering
• Grounding Context: Providing retrieved documents to the LLM
• Relevance Validation: Using LLMs to validate retrieval quality

Related Documentation

• See intrinsics/ for RAG evaluation intrinsics
• See mini_researcher/ for a complete RAG application
• See mellea/stdlib/components/intrinsic/rag.py for RAG helpers