babymore87
MyBoerse.bz Pro Member
Advanced RAG Techniques Architecture [2026]
Published 8/2026
MP4 | Video: h264, 1920x1080 | Audio: AAC, 44.1 KHz, 2 Ch
Language: English | Duration: 3h 55m | Size: 4.76 GB
Semantic chunking, hybrid search, re-ranking, HyDE, Self-RAG, Graph RAG, RAG evaluation, and governance
What you'll learn
Define RAG, decide when you need retrieval, and run a first retrieve-then-generate program on a real document corpus in Python.
Diagnose why fixed chunks lose a company name, keep a golden dataset with canaries, and pick an embedder from a same-chunks bake-off.
Implement semantic chunking and small-to-big retrieval, then keep the method a recall and chunk-size table actually pays for.
Build hybrid search with keyword plus vector retrieval, fuse ranks with Reciprocal Rank Fusion, rerank, and pack against lost-in-the-middle.
Apply query rewriting, HyDE, a Self-RAG retrieve-or-not gate, Adaptive routing, and Corrective RAG with web search off.
Run Graph RAG only after vector RAG fails a global question, then retrieve a table cell and an image caption from a layout-parsed PDF.
Measure faithfulness and cost per query, add a semantic cache and traces, and enforce tenant filters, PII redaction, and an audit row.
Requirements
Comfort reading and writing Python in VS Code (functions, virtualenv, pytest).
Python 3.11 or later, Git, and about 3 GB of unused disk.
VS Code. An API key is not required for retrieve or the test suite. Generate can use Qwen or a local Ollama model.
No prior RAG course is required. The opening sections build the naive baseline. You do need to read a metrics file.
Description
This course contains the use of artificial intelligence.
This Advanced RAG Techniques course takes you from a first retrieve-then-generate program to a measured production RAG pipeline on one Python workbench.
You chunk documents, embed them, take top-k, and stuff the prompt. The system answers. Then a ticket asks what a company earned last quarter, and the retrieved sentence is true and missing the company name. That is an accuracy ceiling, not a model failure.
You do not need a prior RAG course. We start with what RAG is in plain words, then you build a first working program, then you measure every upgrade on the same corpus and the same question file. Engineers already running naive RAG can skip ahead after the baseline.
Anthropic described that identity failure in Introducing Contextual Retrieval on 19 September 2024. On their mix, contextual embeddings plus BM25 plus a reranker cut top-20 retrieval failure 67 percent (5.7 percent to 1.9 percent). Those are their numbers, dated September 2024, not a promise about your documents. This course teaches you to run the same kind of keep-or-kill test on yours.
Build a Measured Advanced RAG Pipeline on One Workbench
What you will be able to do
- Define RAG, decide when you need retrieval versus a long context, and run a first retrieve-then-generate program in VS Code.
- Point one client at Qwen, any OpenAI-compatible provider, or a local Ollama or LM Studio server.
- Implement semantic chunking and small-to-big retrieval, then keep the method a recall table actually pays for.
- Build hybrid search with keyword plus vector retrieval, Reciprocal Rank Fusion, cross-encoder reranking, and prompt packing.
- Apply query rewriting, HyDE, a Self-RAG retrieve-or-not gate, Adaptive routing, and Corrective RAG with web search off.
- Run Graph RAG only after vector RAG fails a global question, then retrieve a table cell and an image caption.
- Measure faithfulness and cost per query, add a semantic cache and traces, and enforce tenant filters, PII redaction, and an audit row.
Who this course is for
AI engineers, ML engineers, and software developers who will ship retrieval over private documents. You can start with zero RAG knowledge if you already write Python. This is not a no-code course.
What is inside this course
The course is 24 teaching sections, 137 lectures, about 13 hours of video. Each section ends in something you can run or defend at a keyboard.
Start from zero RAG
1. What RAG is, when you need it, the five pipeline stages, and how a program calls the workbench.
1B. How a language model answers, why fluent text can be invented, and why RAG passes your files through that desk.
2. Point the same client at Qwen, any OpenAI-compatible provider, or a local Ollama or LM Studio server, and run a live generate ping.
2B. Embed two real sentences, read a cosine table, and print one chunk from a filing.
2C. Store chunks, ask a neighbor question, and walk mini_rag .py line by line. That is your first RAG program.
Naive baseline, then the failure
3. Run naive chunk, embed, retrieve, then generate on the sample corpus. Persist the same chunks to Chroma, FAISS, Qdrant, or optional pgvector.
4. Show why a true sentence without its document header fails an identity ticket. Keep a tagged golden file with canaries. Pick an embedder from a same-chunks bake-off.
Retrieval upgrades you can measure
5. Swap chunkers on the same corpus, including a cosine-breakpoint semantic splitter, and pick one from recall and chunk size.
6. Run parent-child, sentence-window, and auto-merging retrieval. Try late chunking once. Keep only the method the board pays for.
7. Run BM25 and dense in parallel, fuse ranks with Reciprocal Rank Fusion, and recover a token naive dense missed.
8. Prepend chunk context, rerank a shortlist with a cross-encoder, pack winners at the ends of the prompt, and show a recall at k lift.
8B. Write a two-second rerank budget and escalate whole documents when retrieve says unanswerable.
9. Turn one user question into better searches with rewrite, multi-query, and HyDE. Drop any rewrite whose cost exceeds its recall lift.
Loops, graphs, and documents that are not plain text
10. Run a Self-RAG retrieve-or-not gate, refuse an unsupported answer, and route mixed traffic to none, single, multi, or a named source. We ship a prompt loop and say so.
11. Grade a retrieved set as Correct, Incorrect, or Ambiguous. Treat retrieved text as data. Run a retrieve, grade, rewrite or answer loop with web search off.
11B. Cap the retrieve loop and compose two hops without dropping the company token.
12. Name a global question, show vector RAG fail it, run a small Graph RAG community summary, and refuse a full graph when local questions hold.
13. Smash a PDF table with a naive text split, restore the rows with a layout parser, and retrieve a table cell plus an image caption.
Eval, cost, governance, then ship
14. Run faithfulness and context recall on two pipeline versions and keep the winner from metrics.json.
15. Print cost per query, serve a repeated question from a semantic cache without generating, and write an OpenTelemetry-shaped trace.
15B. Upsert one changed file, tombstone the old ids, and refuse a mixed embedder space.
16. Pre-filter retrieval by tenant, redact a PII field, export an audit row of chunks the model saw, and name ingest poisoning plus EchoLeak.
17. Walk a RAG design in interview order and defend the stack you kept with a metrics file and a one-page decision note.
18. Optional. Name five 2026 frontier cards and refuse each as a Monday default. Skip this if you are shipping the measured stack.
How this course runs
This is a professional Advanced RAG Techniques tutorial for developers in VS Code and Python. It is not a LangChain certification, a paper seminar, or a no-code tour. If you want a RAG tutorial that stays a chatbot, this is the wrong listing. If you want an advanced RAG architecture you can defend to a staff engineer, this is the workbench.
How you learn: each concept is a picture, then lab parts in the highlighter, then a full run that keeps or kills the technique. Numbers on screen come from the runner on this corpus, or they are quoted as someone else's dated result.
What is included: 137 lectures, about 13 hours of video, 62 concept lessons and 73 screen walkthroughs. A knowledge check ends most concept lectures. A quiz closes most teaching sections. Three coding exercises cover a chunker, Reciprocal Rank Fusion, and faithfulness. Four assignments add domain questions, defend a rewrite, write local versus global questions, and ship a decision note. A timed practice test is tagged by pillar so a wrong answer points you back to what to restudy.
The companion project is provided with an original ACME policy pack.
What you need: Python 3.11 or later, Git, VS Code, about 3 GB of disk. An API key is not required to run retrieve or tests. Generate can use Qwen or a local Ollama model. You do not need a prior RAG course. Last updated August 2026.
Who teaches this course?
Pragati Kunwer (he/him) has taught 19,152 students on his AI Agents Bootcamp, counted on 15 August 2026. His instructor profile listed 32,566 total learners and 2,240 reviews on 14 August 2026. He has spent over 20 years shipping AI for Fortune 500 companies including IBM and Nike. He built this Advanced RAG Techniques course because capable enterprise teams kept collecting method names and still shipping the naive loop. He has sat in rooms where a retrieved chunk was true and useless, and where Graph RAG was proposed as the default index for a FAQ. This course is the workbench he wishes those teams had opened first: one corpus, one question file, every change measured, and a written rule for when to refuse.
Who this course is for
AI engineers and ML engineers who will ship retrieval over private enterprise documents.
Software developers adding search-plus-generate to a product who need hybrid search, evaluation, and access control.
Developers who finished a first RAG tutorial and need the measured sequel, not another chatbot tour.
This is not a no-code course. If you never write Python, this is not for you.
Homepage
Recommend Download Link Hight Speed | Please Say Thanks Keep Topic Live
No Password - Links are Interchangeable
