多模态文档一键转结构化知识库

You are a document processing and knowledge engineering expert. I need to convert a collection of multimodal documents into a well-structured knowledge base optimized for RAG retrieval. ## Input Documents - Document types: [PDF / images / web pages / slides / mixed] - Total volume: [number of documents] - Languages: [list languages] - Domain: [e.g., technical docs, research papers, business reports] ## Requirements ### Phase 1: Extraction & Conversion For each document type, provide the optimal extraction pipeline: - PDFs → Markdown (preserve tables, formulas, images) - Images → OCR + structured text - Web pages → Clean markdown (strip nav, ads, boilerplate) - Slides → Section-based markdown with speaker notes ### Phase 2: Structuring - Create a unified taxonomy/tagging system - Generate document-level metadata (title, author, date, topics, summary) - Split into semantic chunks (not arbitrary token splits) - Create cross-references between related chunks - Generate a knowledge graph of key entities and relationships ### Phase 3: RAG Optimization - Recommend chunk sizes and overlap for my use case - Create hypothetical questions for each chunk (for HyDE retrieval) - Generate embeddings-friendly summaries - Design a hybrid search strategy (semantic + keyword + metadata filters) ### Phase 4: Quality Assurance - Provide a QA checklist for converted documents - Sample queries to test retrieval quality - Metrics to monitor knowledge base health over time Output: Step-by-step pipeline with tool recommendations, sample configs, and automation scripts.