← Back to Case Studies
Automation & AI Engineering

Document Intelligence: Retrieval-Augmented Generation Over Large PDF Collections

Built a RAG platform that turns large PDF collections into a queryable knowledge base — semantic retrieval over PostgreSQL and pgvector, page-level citations behind every generated claim, and structured report export to Word and PDF.

RAG

Retrieval-augmented generation

Page-level

Citation tracking

pgvector

Semantic search at scale

Multilingual

Including right-to-left scripts

The Challenge

Organisations sitting on hundreds or thousands of PDFs generally cannot search them in any useful way:

  • Collections run to hundreds of documents — far too many to read manually under any deadline
  • Keyword search misses anything phrased differently to the query; full-text search returns whole documents when the answer was one paragraph
  • Every generated claim needed a verifiable, page-level citation back to its source document
  • Separate workspaces needed genuine data isolation, with no cross-workspace retrieval
  • Document sets spanned multiple languages, including right-to-left scripts, without a separate pipeline per language
  • Output had to leave the system as Word and PDF documents, not as chat transcripts

The Solution

I designed and built a two-pipeline RAG system — ingestion runs once per document, query runs per question — so re-indexing and model changes stay independent of each other:

Layout-Aware Ingestion

PDF extraction that preserves multi-column layouts and tables, with chunk boundaries tuned so splits do not land mid-table or mid-clause.

Vector Search & Retrieval

Chunks are embedded and stored in PostgreSQL with pgvector, retrieved by semantic similarity with reranking applied before anything reaches the model.

Citation-Preserving Generation

Page provenance is attached at ingestion and carried through retrieval, so citations are a property of the pipeline rather than something the model is asked to produce.

Structured Report Output

LlamaIndex orchestrates GPT-4 to produce a fixed report shape rather than free-form answers, making results comparable across queries.

Workspace Isolation

Multi-tenant by design — each workspace has its own document set and index, with retrieval scoped so nothing crosses the boundary.

Export & Distribution

Reports export to Word and PDF with formatting and citations intact, ready to circulate without reformatting.

Results

The platform turns an unsearchable document pile into a system that answers questions with evidence attached:

Hours → minutes

Query hundreds of documents and receive a structured report in minutes

Verifiable output

Every claim links back to a specific page in a specific source document

Model-portable

Generation routed through a LiteLLM proxy, so swapping models is a config change

Multilingual

One pipeline handles document sets spanning several languages and scripts

Technology Stack

Frontend

  • • React

Backend

  • • FastAPI
  • • Python

AI & ML

  • • GPT-4
  • • LlamaIndex
  • • Haystack 2.x
  • • LiteLLM

Database & Infrastructure

  • • PostgreSQL
  • • pgvector
  • • Docker

Need something similar?

I build custom platforms, automation systems, and data solutions. Let's discuss your project.