DocVault - Intelligent Document RAG with Exact Citation

# DocVault - Intelligent Document RAG with Exact Citation

## Autonomous Execution Instructions

You are running autonomously via Ralph. DO NOT ask questions or wait for user input. NEVER present options or ask "What would you like me to work on?". Instead:

1. Read `fix_plan.md` to find the next unchecked task (`- [ ]`)
2. Read the relevant spec in `specs/` for detailed requirements
3. Read `CLAUDE.md` for conventions and project structure
4. Implement the task completely (code + tests)
5. Mark the task as done in `fix_plan.md` (`- [x]`)
6. Move to the next task

If a task is unclear, make a reasonable decision based on the specs and proceed. If a task is blocked by a dependency, skip to the next unblocked task.

Work through tasks **in order**, top to bottom. Each loop iteration should complete at least one task.

## Project Overview

DocVault is a production-grade RAG (Retrieval-Augmented Generation) system that answers questions from documents with **exact citations** — highlighting the precise passage in the original document. It supports PDFs (scanned and digital), DOCX, PPTX, and images.

The system is designed to showcase AI Engineering best practices: evaluation pipelines, agentic retrieval, guardrails, and observability.

## Goals

1. **Exact Citation RAG**: Answer questions and cite the verbatim source passage with page number and bounding box coordinates
2. **Multi-Format Support**: Handle digital PDFs, scanned documents (OCR), DOCX, PPTX, and images
3. **Provider-Agnostic LLM**: Use LiteLLM to support Gemini, Claude, GPT, Ollama, etc. via a single interface
4. **Evaluation-Driven**: Every component has automated metrics, not just "it looks right"
5. **Production Mindset**: Guardrails, observability, and security from day one

## Architecture