Loading...
Loading...
Part of the accelerator
Vectorless RAG with Hierarchical Document Trees is one workshop inside AI Engineer Career Accelerator Program.
See the full path with mentor reviews and a portfolio outcome.
Parse PDFs into a hierarchical document tree. Let a LangGraph agent walk the structure and reason about which sections answer the question. Ship retrieval with zero vector infrastructure and citations your users can actually verify.
Engineers are learning here from
Retrieve without embeddings using hierarchical document trees and LLM reasoning. Zero vector database, zero infrastructure cost. Build a LangGraph agent that navigates PDFs section by section and returns cited answers.
Skip the vector database. Navigate document trees with an LLM agent and get cited answers.
What you'll ship
What you'll learn
Curriculum
Why vectorless
Understand exactly where vector RAG breaks and why document structure is often the better signal
PDF to DocumentTree
Parse PDFs into a hierarchical TreeNode structure with layout-aware markdown and JSON caching
Section summaries
Attach a short summary to every node so the agent can route without reading full section text
LangGraph retrieval agent
Build a state machine that analyzes nodes, routes conditionally, retrieves content, and generates cited answers
Production polish
Visualize the workflow, extend to multiple documents, and instrument cost and latency
Who it's for
whose chunked pipelines return near matches but never the right section of a multi-page document
who want to ship retrieval without running and paying for a vector database
building question answering over research papers, legal filings, or technical manuals with real structure
who need a principled comparison between vector search and structure-aware retrieval
FAQ
No. Vectorless shines on structured documents with clear hierarchy (papers, manuals, contracts) and small to medium corpora. Vector RAG still wins for large unstructured corpora where semantic similarity matters more than document structure. You will learn the decision criteria.
Neither. The pipeline runs on CPU with PyMuPDF4LLM for parsing and any OpenAI-compatible LLM for reasoning. The only infrastructure is an API key.
Typically one call per node visited during traversal plus one final generation. For a well-structured paper, that lands around four to six calls per question. You will instrument the pipeline and see exact numbers.
Yes. The final module covers multi-document trees where the agent chooses which document to descend into before walking its sections.
Architecture choices compound. Pick the one that matches how your documents are actually written.