“X-RAG” – Extended Retrieval-Augmented Generation for Complex PDF Content Extraction
Complex PDF documents often blend multi-page text with rich visual information, presenting a significant challenge for conventional data extraction. Traditional “text-only” pipelines typically parse each page as an isolated text chunk, overlooking crucial cross-page references and visual cues embedded within images. This gap leads to incomplete or fragmented information retrieval.
To address these limitations, we introduce “X-RAG,” or Extended Retrieval-Augmented Generation. By coupling page-to-image conversion with large language model (LLM) processing, X-RAG ensures the full context—including multi-page text flows and intricate visual details—is captured and transformed into a structured representation (e.g., XML).
X-RAG extends traditional Retrieval-Augmented Generation in two key ways:
1. Input PDF
The process begins with a PDF document.
2. Convert PDF Pages to Images
Each page of the PDF is turned into a corresponding image (often done so that any visual elements can be processed further, for example by an AI model or an OCR system).
3. Batching + Context
Before further processing, the images (and/or their text segments) are grouped or batched. This step also involves organizing relevant context (such as metadata or additional information) needed for downstream tasks. Using the LLM's multi-modal capabilities for Images, we curate the relative context between images, in order to batch images according to each other's relative content.
4. PDF Text Extraction
In parallel (or as a complementary process), the text layer is extracted from the original PDF. This might be done via OCR (if the PDF does not have a searchable text layer) or by simply reading the embedded text if available.
5. LLM Image Enrichment
Using a Large Language Model (LLM), the images are analyzed or “enriched.” This typically means generating descriptions, identifying key elements, or extracting metadata that augments the extracted PDF text with more detailed context or semantic information.
6. XML Assembly
All of the outputs (the extracted text, image descriptions, and any other contextual data) are combined to form an XML structure. This involves mapping the text and image metadata into a specified XML schema or format.
7. XML Repair & Validation
The assembled XML is then checked for correctness. Any structural or syntactic errors are repaired, and the content is validated against a target schema or set of rules to ensure completeness and compliance.
8. Final XML
The process concludes with a properly structured, validated XML file that integrates all the information extracted and enriched from the PDF.
By carrying forward “in-progress” XML from earlier batches, the system avoids disjointed references. Sections spanning multiple pages become a coherent, single narrative.
Enabling the LLM to “see” page images means it can interpret diagrams, call out captions, and even handle textual overlays that standard text extraction might miss.
The final XML preserves hierarchical organization—from high-level <manual> tags, to <section> tags, down to <page>and <image> elements—making it far simpler to index, search, or reuse the data.
X-RAG can scale to very large PDFs by adjusting batch size, ensuring the LLM remains within token limits.
The approach is modular: each batch is processed independently, and final XML pieces are concatenated and cleaned.
LLMs have context window limits. Large documents require careful batching (e.g., 2–3 pages at a time) to manage token usage.
In reality, pages may vary widely in design (column-based text, large images, embedded fonts). A robust image rendering process is essential to capture content reliably.
The system’s final step validates XML structure to prevent malformed tags or partial outputs from earlier incomplete steps.
Long user guides with cross-referenced diagrams. X-RAG ensures the transitions between textual instructions and diagrams are preserved.
Important references can span multiple pages, and images or tables may contain key provisions. X-RAG keeps each reference in context with relevant visual cues.
Multi-page figures (e.g., figure splits across consecutive pages) remain logically coherent. Supporting text is properly tied to each figure element.
X-RAG marries the strengths of large language models with extended retrieval across page-spanning context and images. By preserving visual, textual, and structural continuity in an XML output, X-RAG ensures no critical information is lost, enabling next-level insights and analytics over complex, image-heavy PDF documents.
This method surmounts the limitations of purely text-centric pipelines, offering a robust, holistic view of the document’s information. Whether it’s technical manuals, regulatory filings, or research papers, X-RAG yields complete and cohesive datasets ready for advanced downstream processing.
Use this button to send us an email, we're happy to meet with you and find a solution to your challenge!