01 · Overview
Product context and engineering scope.
The goals, operating context, and technical decisions behind this case study.
Built an end-to-end vectorless Retrieval-Augmented Generation pipeline that processes PDF documents and answers user questions without relying on embeddings, semantic similarity search, or a vector database.
01
PageIndex transforms each document into a hierarchical tree containing section titles, summaries, page references, text content, and parent-child relationships, preserving the original document structure.
02
LangChain and Groq Llama 3.1 perform tree-based retrieval by selecting relevant node IDs, recursively retrieving the corresponding sections, and generating grounded answers with section-level and page-level references.
02 · Capabilities
Capabilities designed around real user needs.
The product features and system behaviors delivered by this project.
01
Hierarchical PDF Indexing
Processes PDF documents with PageIndex and converts them into structured trees containing sections, nested subsections, summaries, text, node IDs, and page references.
02
Vectorless Document Retrieval
Retrieves relevant document content without embeddings, similarity search, or vector databases by reasoning directly over the document hierarchy.
03
LLM-Based Tree Search
Uses Groq Llama 3.1 to analyze a compressed document tree and return the node IDs most relevant to the user query.
04
Recursive Node Retrieval
Traverses nested PageIndex nodes recursively and retrieves the complete content of sections selected during the tree-search stage.
05
Grounded Answer Generation
Generates answers strictly from retrieved document sections, reducing unsupported responses and keeping outputs grounded in source content.
06
Section & Page References
Returns supporting section titles and page numbers so users can trace generated answers back to the original document.
07
Asynchronous Document Processing
Submits documents to PageIndex, monitors processing status, and retrieves the completed document tree after indexing finishes.
08
Multi-Query Evaluation
Supports testing multiple natural-language questions against the same indexed document while displaying selected nodes and retrieved sections.
03 · Decisions
Constraints translated into engineering decisions.
The problems that shaped the product and the responses used to address them.
01
The constraint
Retrieving relevant context without embeddings or vector similarity search
Engineering decision
Compressed the hierarchical PageIndex tree and used an LLM-based selection step to identify relevant node IDs based on section titles, summaries, and document structure.
02
The constraint
Preserving document structure and source traceability during retrieval
Engineering decision
Maintained node IDs, nested section relationships, titles, page indexes, and full section text throughout tree traversal and answer generation.
03
The constraint
Preventing the language model from generating unsupported answers
Engineering decision
Restricted answer generation to retrieved node content and instructed the model to reference the corresponding section titles and page numbers.
04 · Technology
Technology and tools
The complete technology stack utilized in this project, organized by engineering area.
More selected work
Continue exploring the portfolio case studies.