Ch 5: RAG — Under the Hood

Ch 5 — RAG Under the Hood

Inside text splitters, embedding math, FAISS indexes, retriever internals, and the LCEL RAG pipeline

Index ← High Level

Under the Hood

Click play or press Space to begin...

Step- / 10

AText Splitter InternalsHow RecursiveCharacterTextSplitter actually splits

description

Raw TextFull document
as a string

split at

list

Separators\n\n → \n → . → " "
recursive fallback

merge

view_column

Chunkschunk_size limit
+ overlap window

arrow_downward Embedding model internals: text → dense vector

BEmbedding Model InternalsTokenize → transformer → pool → normalize

text_fields

Input Text"refund policy
for customers"

tokenize

data_array

Token IDs[8134, 4813,
369, 6890]

encode

neurology

TransformerContextual token
representations

pool

scatter_plot

Dense Vector1536-d normalized
unit vector

arrow_downward Vector store internals: FAISS index structure

CVector Store & FAISS InternalsHow vectors are indexed and searched efficiently

scatter_plot

VectorsN vectors ×
1536 dimensions

index

account_tree

FAISS IndexIndexFlatIP or
IndexIVFFlat

sort

Top-k ResultsDistances + indices
→ Document objects

arrow_downward Retriever as a Runnable: the VectorStoreRetriever class

DRetriever Runnable InternalsVectorStoreRetriever wraps the store as a Runnable

text_fields

Query String"refund policy"
raw user input

.invoke()

manage_search

Retriever_get_relevant_
documents()

delegates

database

Vector Storesimilarity_search
(query, k=4)

returns

article

List[Document]page_content +
metadata

arrow_downward LCEL RAG chain: RunnableParallel and data flow

ELCEL RAG Chain InternalsRunnableParallel branches, then sequential pipeline

call_split

RunnableParallelcontext + question
run concurrently

merge

data_object

Dict Output{context: str,
question: str}

pipe

linear_scale

Sequentialprompt → model
→ parser

arrow_downward Advanced: MMR, cross-encoder reranking, parent document retriever

FAdvanced Retrieval InternalsMMR diversity, cross-encoder scoring, parent-child chunks

hub

MMRRelevance minus
redundancy score

compare_arrows

Cross-Encoder[query, doc] →
single score

account_tree

Parent DocSmall chunks search
→ return full parent