A Full RAG Chain

# ===== Pipeline overview =====
# docs -> chunks -> embeddings -> vector store
# question -> retrieve -> rerank -> prompt -> LLM -> answer + citations
# offline: evaluate retrieval + answer faithfulness

# ===== Ingestion =====
from langchain_community.document_loaders import DirectoryLoader, TextLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores import Chroma

loader = DirectoryLoader('./corpus', glob='**/*.md', loader_cls=TextLoader)
docs   = loader.load()

splitter = RecursiveCharacterTextSplitter(
    chunk_size=800, chunk_overlap=120,
    separators=['\n\n', '\n', '. ', ' ', ''],
)
chunks = splitter.split_documents(docs)

emb = OpenAIEmbeddings(model='text-embedding-3-small')
store = Chroma.from_documents(chunks, emb, persist_directory='./chroma')
store.persist()

# ===== Retrieve + rerank =====
from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import CohereRerank

base = store.as_retriever(search_kwargs={'k': 20})
rerank = CohereRerank(model='rerank-english-v3.0', top_n=5)
retriever = ContextualCompressionRetriever(
    base_compressor=rerank, base_retriever=base
)

# ===== Prompt + LLM =====
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser

PROMPT = ChatPromptTemplate.from_template('''
You are a careful assistant. Answer ONLY from the provided context.
If the context does not contain the answer, say so explicitly.

Cite sources as [#] using the source labels.

# Question
{question}

# Context
{context}
'''.strip())

def format_context(docs):
    return '\n\n'.join(f'[{i+1}] {d.metadata.get("source","?")}\n{d.page_content}' for i, d in enumerate(docs))

llm = ChatOpenAI(model='gpt-4o-mini', temperature=0)

chain = (
    {'context': retriever | format_context, 'question': RunnablePassthrough()}
    | PROMPT
    | llm
    | StrOutputParser()
)

print(chain.invoke('What is our SLA for incident response?'))

# ===== Evaluation =====
# Build a small golden set (50-200 Q+A pairs).
# Score on two axes:
#   1. Retrieval recall  - is the right chunk in top-k?
#   2. Answer faithfulness - does the answer match the cited chunk?
from langchain.evaluation import load_evaluator
faithfulness = load_evaluator('labeled_pairwise_string', criteria='faithfulness')
# loop golden set, call chain, score, log to a CSV.

# ===== Patterns to internalise =====
# - Chunk on paragraph boundaries, not fixed tokens
# - Always rerank when you have > 5 candidates
# - Prompt the model to refuse rather than hallucinate
# - Cite by chunk index, render label in UI
# - Evaluate retrieval and answer separately; they fail differently

# ===== Pitfalls =====
# - chunk_size too small  -> chopped sentences ruin embeddings
# - chunk_size too large  -> low precision on retrieval
# - Cosine vs dot product mismatch between embedder and store
# - No metadata filter on multi-tenant corpora -> data leaks
# - Trusting answers without faithfulness eval -> silently confident wrong answers

from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser

rag_prompt = ChatPromptTemplate.from_template('''Answer using only the context.\n\nContext: {context}\n\nQuestion: {q}''')

rag = ({'context': retriever, 'q': RunnablePassthrough()} | rag_prompt | llm | StrOutputParser())
print(rag.invoke('How do I get a refund?'))