Advanced RAG Bootcamp

Advanced RAG extends a foundational RAG pipeline into enterprise retrieval: complex and multimodal documents, tables, multi-source queries, changing corpora, and the question of when graph structure is warranted. A baseline pipeline gets additive layers rather than a rewrite, and each method choice is weighed against retrieval quality, grounding, cost, latency, and engineering effort.

The repository pairs a shared library (aieng-advanced-rag) with self-contained reference implementations, each a worked example of a distinct enterprise RAG problem and the techniques that fit it. Enter at whichever one is closest to the target system.

The repository is the course material of Vector Institute's Advanced RAG Bootcamp, published for anyone who builds or studies retrieval systems: engineers taking a RAG pipeline past the baseline, and researchers comparing Advanced RAG with GraphRAG on shared data. You don't need a bootcamp account or a model key to follow it. Every notebook ships with RUN_EXPENSIVE = False, and in that mode it replays the committed results instead of calling a model. Run without bootcamp keys explains how, and how to point the notebooks at your own endpoint.

First time here? Run implementations/getting_started/00_environment_check.ipynb before anything else. It confirms that the dependencies, keys, embedding model, and data are in place, and names the fix for anything that is not. docs/setup.md covers setup.

Teaching model: one spine, two circles

The bootcamp is organized along one axis, RAG becoming agentic. RAG basics are assumed.

Baseline RAG → Advanced RAG (hybrid, rerank, contextual/hierarchical chunking, query rewriting) → Agentic RAG (ReAct loops, decomposition, self-correction, reflexion) → Agent-native / embedding-less retrieval (the agent is the retriever: grep/navigate/tool-call, Docling-agent- and coding-agent-style, no vector index in the loop).

Two overlapping circles, Retrieval (the "R") and Agent (the loop reasoning over it). Along the spine the Agent circle grows and slides over the Retrieval circle until, at the terminal, the two coincide and the embedding index disappears. Quality moves by turning either circle. A richer retrieval substrate with a simple agent (KG-RAG + basic ReAct), or a plain substrate with a deep agentic loop, can reach comparable quality, so the portfolio holds one constant and varies the other to read the cost, latency, and complexity trade-off directly.

The Advanced-RAG-vs-GraphRAG decision is applied at each stage of the spine. GraphRAG stacks on the Advanced RAG spine, warranted by specific capability gaps (entity aggregation, explicit reasoning paths, corpus-scale summarization, path-finding multi-hop). Everything sits downstream of a shared evaluation harness.

Reference implementations

Each reference implementation (RI) is independent and self-contained; see its README.md for the full walkthrough. Together they span the spine from a conventional baseline to agent-native, embedding-less retrieval, and each pairs an Advanced RAG method with a GraphRAG contrast. Each one's CORPUS.md or README.md records its datasets, their licences, and their pinned versions. The circles under each folder name the retrieval elements that RI combines.

#	Implementation	Spine position	The problem
1	regulatory_policy_qa/	Baseline → Advanced RAG	The Advanced RAG spine and shared evaluation harness, with reliable "not stated" null-answer detection on dynamic regulatory content.
2	cross_document_synthesis/	Advanced → Agentic RAG	Flagship. Multi-hop reasoning across many documents, growing the agent loop, with the Advanced-RAG-vs-GraphRAG decision resolved on measured results.
3	table_structured_reasoning/	Advanced RAG (substrate)	Retrieval and answering over tables and structured data (Docling section folders, table-aware chunking), grounded on financial filings.
4	graphrag_technical_support/	Agentic RAG over a graph	The full KG-RAG track: graph construction, hybrid vector+graph retrieval with provenance, agentic traversal, and graph maintenance.
5	agent_native_retrieval/	Agent ∩ Retrieval overlap	The spine terminal: embedding-less, agent-native retrieval (grep/navigate/Docling-agent), benchmarked against the vector baseline on the same corpora.

Design notes

Every RI starts from the same baseline pipeline and evaluation harness, and each advanced technique is an added layer on top of it. Techniques are ordered by expected return: hybrid search, reranking, and hierarchical chunking come before contextual retrieval, then adaptive routing, then agents, then graphs. Each RI holds either the retrieval substrate or the agent loop fixed and varies the other, so a change in cost, latency, or quality traces to one circle. Each RI also shows where its technique loses, and ends with a starter-agent notebook that moves the use case one stage further along the spine. Datasets are public and come with questions and answers, so lab time goes to retrieval rather than data preparation.

Repository structure

aieng-advanced-rag/  # Shared library for the bootcamp
implementations/     # Environment check, self-contained reference implementations, and special topics
docs/                # Setup guide and additional documentation
scripts/             # Workspace setup and data scripts
tests/               # Onboarding gate that the platform runs (not run in CI)

implementations/special_topics/ holds short, standalone studies that apply to more than one RI: citations (scoring cited answers), synthetic_question_sets (building an evaluation set from an unlabelled corpus), and visual_documents (parsing, enriching, or embedding the page).

Getting started

On a Coder workspace, scripts/setup.sh runs on every start. It installs the dependencies, loads your bootcamp keys into the shell, fetches the core data, and starts Jupyter. On your own machine, follow Set up on your own machine.

1. Run 00_environment_check.ipynb and fix anything it reports.
2. Begin with regulatory_policy_qa/ to learn the pipeline and evaluation harness.
3. Fetch the data for the RI closest to your use case, then move to it: python -m aieng.advanced_rag.corpora.download <ri_number>.

To work on your own documents, start from RI #1's 04_bring_your_own_corpus.ipynb. It reads a folder of .txt, .md, or .pdf files, generates a question set with known answers and null questions, and ranks the retrieval arms on it. 07_levers_and_answers_on_your_corpus.ipynb scores the retrieval levers and answering with abstention on that set. 99_starter_agent.ipynb runs an agent over the same folder. Share a team's work covers a team fork and merging repository updates. If your data can't come onto the platform, Run the notebooks on data that stays off the platform covers your own environment and endpoint.

Bootcamp keys (OPENAI_API_KEY, OPENAI_BASE_URL, and LANGFUSE_*) live in your shell environment. Personal keys go in .env, as .env.example lists. docs/setup.md explains the credential model and has a troubleshooting table.

Run without bootcamp keys

You can read and run every notebook without a key. Each notebook ships with RUN_EXPENSIVE = False, and in that mode it replays the committed results in the RI's outputs/ folder through the shared cache instead of calling a model. RIs whose results are too large to commit name a manifest, reference-artifacts.json, that points at files on this repository's GitHub releases. python -m aieng.advanced_rag.cache <manifest> downloads them and checks each digest, and Install the reference artifacts has the details. Embeddings are computed locally, so retrieval runs without a key too.

To make live calls with your own provider key, point the package at any OpenAI-compatible endpoint. In .env, set OPENAI_BASE_URL to the endpoint and VECTOR_PROXY_API_KEY to your key. The package reads both through aieng.advanced_rag.proxy, and an OPENAI_API_KEY exported in the shell is ignored unless OPENAI_BASE_URL points at the Vector proxy, so that an OpenAI key never reaches the proxy by mistake. Then name models your endpoint serves, either with choose_models(generation=..., judging=..., extraction=...) or with the variables ADVANCED_RAG_GENERATION_MODEL, ADVANCED_RAG_JUDGE_MODEL, and ADVANCED_RAG_EXTRACTION_MODEL. choose_models accepts a model outside its default sets with a warning. The default sets use the Vector proxy's model names, and the paid families share gemini-3.8-flash as the judge, so scores you produce with another judge are comparable only with each other. The RI #4 graph build scripts read LLM_BASE_URL and LLM_API_KEY instead.

License

This project is licensed under the terms of the LICENSE file.

Contribution

See our CONTRIBUTING.md guide.

Contact Information

For more information, contact the Vector AI Engineering team at advanced-rag@vectorinstitute.ai. Report a security vulnerability privately, as SECURITY.md describes.