RAG Patterns
Expert guidance for Retrieval-Augmented Generation systems.
Basic RAG Pipeline
python
from openai import OpenAI
from sentence_transformers import SentenceTransformer
import chromadb
# Initialize
client = OpenAI()
embedder = SentenceTransformer('all-MiniLM-L6-v2')
chroma = chromadb.Client()
collection = chroma.create_collection("documents")
# Index documents
def index_documents(documents: list[str]):
embeddings = embedder.encode(documents)
collection.add(
documents=documents,
embeddings=embeddings.tolist(),
ids=[f"doc_{i}" for i in range(len(documents))]
)
# Retrieve and generate
def rag_query(query: str, top_k: int = 5) -> str:
# Embed query
query_embedding = embedder.encode([query])[0]
# Retrieve
results = collection.query(
query_embeddings=[query_embedding.tolist()],
n_results=top_k
)
# Format context
context = "\n\n".join(results['documents'][0])
# Generate
response = client.chat.completions.create(
model="gpt-4o",
messages=[
{"role": "system", "content": f"Answer based on context:\n\n{context}"},
{"role": "user", "content": query}
]
)
return response.choices[0].message.content
Chunking Strategies
Fixed Size Chunking
python
def chunk_fixed_size(text: str, chunk_size: int = 500, overlap: int = 50) -> list[str]:
chunks = []
start = 0
while start < len(text):
end = start + chunk_size
chunk = text[start:end]
chunks.append(chunk)
start = end - overlap
return chunks
Semantic Chunking
python
from langchain.text_splitter import RecursiveCharacterTextSplitter
splitter = RecursiveCharacterTextSplitter(
chunk_size=1000,
chunk_overlap=200,
separators=["\n\n", "\n", ". ", " ", ""]
)
chunks = splitter.split_text(document)
Sentence-Based Chunking
python
import nltk
nltk.download('punkt')
def chunk_by_sentences(text: str, sentences_per_chunk: int = 5) -> list[str]:
sentences = nltk.sent_tokenize(text)
chunks = []
for i in range(0, len(sentences), sentences_per_chunk):
chunk = " ".join(sentences[i:i + sentences_per_chunk])
chunks.append(chunk)
return chunks
Document Structure Chunking
python
import re
def chunk_by_sections(text: str) -> list[dict]:
# Split by headers
sections = re.split(r'(#{1,6}\s+.+)', text)
chunks = []
current_header = ""
for i, section in enumerate(sections):
if re.match(r'#{1,6}\s+', section):
current_header = section.strip()
elif section.strip():
chunks.append({
"header": current_header,
"content": section.strip()
})
return chunks
Embedding Strategies
Hybrid Embeddings
python
from sentence_transformers import SentenceTransformer
from rank_bm25 import BM25Okapi
class HybridRetriever:
def __init__(self, documents: list[str]):
self.documents = documents
self.embedder = SentenceTransformer('all-MiniLM-L6-v2')
self.embeddings = self.embedder.encode(documents)
# BM25 for keyword matching
tokenized = [doc.lower().split() for doc in documents]
self.bm25 = BM25Okapi(tokenized)
def search(self, query: str, top_k: int = 5, alpha: float = 0.5) -> list[str]:
# Semantic search
query_emb = self.embedder.encode([query])[0]
semantic_scores = np.dot(self.embeddings, query_emb)
# BM25 search
bm25_scores = self.bm25.get_scores(query.lower().split())
# Normalize and combine
semantic_norm = (semantic_scores - semantic_scores.min()) / (semantic_scores.max() - semantic_scores.min())
bm25_norm = (bm25_scores - bm25_scores.min()) / (bm25_scores.max() - bm25_scores.min() + 1e-6)
combined = alpha * semantic_norm + (1 - alpha) * bm25_norm
top_indices = np.argsort(combined)[-top_k:][::-1]
return [self.documents[i] for i in top_indices]
Multi-Vector Embeddings
python
from openai import OpenAI
def create_multi_vector_embeddings(document: str) -> dict:
client = OpenAI()
# Generate summary
summary = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": f"Summarize:\n{document}"}]
).choices[0].message.content
# Generate questions
questions = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": f"Generate 3 questions this answers:\n{document}"}]
).choices[0].message.content
# Embed all representations
embeddings = client.embeddings.create(
model="text-embedding-3-small",
input=[document, summary, questions]
)
return {
"document": document,
"document_embedding": embeddings.data[0].embedding,
"summary_embedding": embeddings.data[1].embedding,
"questions_embedding": embeddings.data[2].embedding
}
Retrieval Strategies
Query Expansion
python
def expand_query(query: str) -> list[str]:
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{
"role": "user",
"content": f"Generate 3 alternative phrasings of this query:\n{query}"
}]
)
alternatives = response.choices[0].message.content.split('\n')
return [query] + [alt.strip() for alt in alternatives if alt.strip()]
HyDE (Hypothetical Document Embeddings)
python
def hyde_retrieval(query: str, retriever) -> list[str]:
client = OpenAI()
# Generate hypothetical answer
response = client.chat.completions.create(
model="gpt-4o",
messages=[{
"role": "user",
"content": f"Write a detailed answer to: {query}"
}]
)
hypothetical_doc = response.choices[0].message.content
# Retrieve using hypothetical document
return retriever.search(hypothetical_doc)
Reranking
Cross-Encoder Reranking
python
from sentence_transformers import CrossEncoder
reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
def rerank(query: str, documents: list[str], top_k: int = 3) -> list[str]:
pairs = [[query, doc] for doc in documents]
scores = reranker.predict(pairs)
ranked_indices = np.argsort(scores)[::-1][:top_k]
return [documents[i] for i in ranked_indices]
Cohere Rerank
python
import cohere
co = cohere.Client("your-api-key")
def cohere_rerank(query: str, documents: list[str], top_k: int = 3) -> list[str]:
results = co.rerank(
model="rerank-english-v3.0",
query=query,
documents=documents,
top_n=top_k
)
return [documents[r.index] for r in results.results]
Generation Strategies
Context Compression
python
def compress_context(query: str, documents: list[str], max_tokens: int = 2000) -> str:
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{
"role": "user",
"content": f"""Extract only the parts relevant to the query.
Query: {query}
Documents:
{chr(10).join(documents)}
Relevant excerpts:"""
}]
)
return response.choices[0].message.content
Citation Generation
python
def generate_with_citations(query: str, documents: list[dict]) -> str:
# Format with source IDs
context = "\n\n".join([
f"[{i+1}] {doc['content']}"
for i, doc in enumerate(documents)
])
response = client.chat.completions.create(
model="gpt-4o",
messages=[
{"role": "system", "content": f"""Answer using the context below.
Cite sources using [1], [2], etc.
Context:
{context}"""},
{"role": "user", "content": query}
]
)
return response.choices[0].message.content
Evaluation
python
from ragas import evaluate
from ragas.metrics import faithfulness, answer_relevancy, context_precision
# Prepare dataset
dataset = {
"question": ["What is X?"],
"answer": ["X is ..."],
"contexts": [["Context 1", "Context 2"]],
"ground_truth": ["X is actually ..."]
}
# Evaluate
results = evaluate(dataset, metrics=[faithfulness, answer_relevancy, context_precision])
print(results)