Retrieval-Augmented Generation (RAG) Practice
Learn how to build production Retrieval-Augmented Generation with LangChain - semantic search over vector indexes, ingesting real documents into clean chunks, and wiring retrieval into an LLM for grounded, cited answers.
Python 3.13Embeddings & Vector Search
Learn how to turn text into vectors, search by meaning with cosine similarity and FAISS, fuse dense and keyword results, rerank for precision, and ship a search service that meets its latency SLOs.
Embedding Space
Embed texts, measure meaning with cosine similarity, and return the true nearest neighbors.
Index And Distance
Configure the vector store's distance to match the embedding model and make recall real.
Hybrid Fusion
Fuse dense and keyword rankings so exact identifiers and paraphrases both find their documents.
Reranker Call
Rerank the retrieved candidates so the right document climbs from position twenty to the top.
Latency Budget
Make a correct search index fast enough for its latency budget - build once, answer in milliseconds.
Search Service SLO
Wire index, filter, and reranker into one search service that meets every SLO in a single run.
Document Ingestion & Parsing
Learn how to turn messy real-world files into a clean, searchable corpus - extract ordered text from PDFs and HTML, keep tables meaningful, strip boilerplate, collapse duplicates, capture metadata, and OCR scanned pages into one trustworthy pipeline.
PDF And HTML Extraction
Extract clean, ordered text from PDFs and HTML into LangChain Documents - columns in reading order, scripts and styles gone.
Table Handling
Turn HTML tables into text that keeps every value attached to its header - merged cells, spans, and boundaries intact.
Boilerplate And Dedup
Strip the cookie banners and collapse the nine copies of the same press release - without ever merging two real documents.
Metadata Extraction
Pull titles, dates, and sections into Document.metadata - canonical formats, correct spans, and honest nulls.
Scanned Page Extraction
Route scans to OCR and born-digital pages past it - recover upside-down pages by confidence and record the provenance.
Junk Drawer Corpus
Run the whole ingestion pipeline on one messy dump - every stage measured, cross-format duplicates collapsed, and search over the result just works.
About this path
This path covers retrieval-augmented generation from both ends: embeddings and vector search, and turning real documents into clean, searchable chunks. You write Python with LangChain, and the challenges are medium difficulty, so you should be comfortable with Python first. No earlier RAG experience is needed. The ingestion module works with PDFs, HTML, tables and scanned pages.
Frequently asked questions
4What will I practice in this path?
2 modules: Embeddings & Vector Search and Document Ingestion & Parsing.
How many challenges are there, and how hard are they?
12 challenges, all medium.
What does this path run on?
Challenges run on Python 3.13.
Do I need to install anything?
No. Every challenge runs in a browser-based IDE with everything already set up, and your code is graded automatically against a test suite.