RAG Cheatsheet: Retrieval-Augmented Generation Pipeline Reference

# Minimum stack — works for >90% of starter projects
pip install langchain langchain-openai langchain-chroma langchain-text-splitters

# Optional reranker + sparse retriever (hybrid)
pip install sentence-transformers cohere rank-bm25

# Provider keys
export OPENAI_API_KEY=sk-...
export COHERE_API_KEY=co-...

# Sanity check the stack
python - <<'PY'
from langchain_openai import OpenAIEmbeddings
emb = OpenAIEmbeddings(model="text-embedding-3-small")
print(len(emb.embed_query("hello")))  # 1536 for 3-small
PY

from langchain_text_splitters import (
    RecursiveCharacterTextSplitter, MarkdownHeaderTextSplitter,
)

# Default: recursive, character-based, hierarchical separators
default = RecursiveCharacterTextSplitter(
    chunk_size=900,            # ~ 200-300 tokens
    chunk_overlap=120,         # ~ 15% overlap
    separators=["\n\n", "\n", ". ", " ", ""],
)
chunks = default.split_text(open("doc.md").read())

# Markdown-aware: keep section context as metadata
md = MarkdownHeaderTextSplitter(headers_to_split_on=[("#", "h1"), ("##", "h2")])
sections = md.split_text(open("doc.md").read())

# Token-based when you must respect a context window precisely
from langchain_text_splitters import TokenTextSplitter
tokens = TokenTextSplitter(chunk_size=350, chunk_overlap=50)

# Code-aware (Python AST-respecting)
from langchain_text_splitters import RecursiveCharacterTextSplitter, Language
py = RecursiveCharacterTextSplitter.from_language(
    language=Language.PYTHON, chunk_size=900, chunk_overlap=80,
)

from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings
from langchain.retrievers import EnsembleRetriever, ContextualCompressionRetriever
from langchain_community.retrievers import BM25Retriever
from langchain.retrievers.document_compressors import CohereRerank

emb = OpenAIEmbeddings(model="text-embedding-3-small")
vec = Chroma(collection_name="docs", embedding_function=emb,
             persist_directory="./chroma")

# 1 · Dense (embedding) + sparse (BM25) retrievers in parallel
dense  = vec.as_retriever(search_kwargs={"k": 10})
sparse = BM25Retriever.from_texts(docs)            # docs = list[str] for BM25
sparse.k = 10

hybrid = EnsembleRetriever(retrievers=[dense, sparse], weights=[0.6, 0.4])

# 2 · Rerank the merged candidates with a cross-encoder
reranker = CohereRerank(model="rerank-english-v3.0", top_n=4)
final = ContextualCompressionRetriever(
    base_compressor=reranker, base_retriever=hybrid,
)

hits = final.invoke("What is the WFH policy?")
for h in hits:
    print(round(h.metadata.get("relevance_score", 0), 3), h.page_content[:80])

# Load -> split -> embed -> retrieve -> rerank -> generate -> cite.
from pathlib import Path
from langchain_community.document_loaders import PyPDFDirectoryLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_chroma import Chroma
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough

docs = PyPDFDirectoryLoader("./pdfs").load()
chunks = RecursiveCharacterTextSplitter(
    chunk_size=900, chunk_overlap=120,
).split_documents(docs)

emb = OpenAIEmbeddings(model="text-embedding-3-small")
store = Chroma.from_documents(chunks, emb, persist_directory="./chroma")
retriever = store.as_retriever(search_kwargs={"k": 6})

PROMPT = ChatPromptTemplate.from_messages([
    ("system",
     "Answer using ONLY the context. Cite sources as [filename:page]. "
     "If unsure, say 'I don't know'.\n\nContext:\n{context}"),
    ("human", "{question}"),
])

def format_ctx(docs):
    return "\n\n".join(
        f"[{d.metadata.get('source')}:{d.metadata.get('page')}] {d.page_content}"
        for d in docs
    )

chain = (
    {"context": retriever | format_ctx, "question": RunnablePassthrough()}
    | PROMPT
    | ChatOpenAI(model="gpt-4o-mini", temperature=0)
    | StrOutputParser()
)
print(chain.invoke("Summarise the access policy in 3 bullets."))