1. The Problem with Classical Vector Chunking
In large-scale enterprise environments, critical standard operating procedures (SOPs), process training manuals, and compliance guidelines are distributed across dense, highly formatted PDF documents containing complex multi-column tables, flowcharts, and cross-referenced clauses.
Traditional RAG architectures rely on semantic text splitting and vector embeddings. When applied to multi-page enterprise SOPs, this naive chunking breaks table rows across arbitrary boundaries, detaches figures from their captions, and causes subtle semantic drift during vector similarity lookups — resulting in hallucinated policy guidelines and untrusted answers.
The Page-Indexed Retrieval Paradigm
VectorlessRather than shredding documents into disconnected vector chunks, the Page-Indexed Architecture treats the entire structured document page as the atomic unit of truth.
- • Chunks arbitrarily split at 500 tokens
- • Tables ripped in half, losing headers
- • High embedding drift on technical jargon
- • Opaque similarity scores without grounding
- • Entire page geometry preserved
- • Complete table & diagram context retained
- • Deterministic section & page routing
- • 100% exact page coordinate citations
2. Technical Stack & Execution Engine
The platform combines LangGraph state machines with the latest OpenAI Responses API and Anthropic Claude 3.5 Sonnet:
- Hierarchical Page Metadata IndexingDocuments undergo layout analysis to extract page-level headers, section hierarchy, table coordinates, and key entities into a fast, searchable document map.
- Intent-Driven Page Router (LangGraph)When an employee asks a process question, the router evaluates candidate page clusters using structured LLM classification, selecting the exact 2-4 authoritative pages.
- Direct Context Synthesis & Citation VerificationThe selected page contents are fed in full to Claude 3.5 / OpenAI Responses API with structured JSON schemas, producing precise answers linked to exact document page numbers.