Overview

Integrate Docling to enhance Archon's document processing capabilities with multi-format support and intelligent chunking for RAG operations.

Why Docling?

• Multi-Format Support: PDF, DOCX, PPTX, XLSX, HTML, Audio (MP3, WAV), Images
• Built-in OCR: No custom OCR implementation required (EasyOCR support)
• Structure Preservation: Maintains tables, sections, hierarchies automatically
• RAG-Optimized: Hybrid chunking strategy respects semantic boundaries
• Unified Output: All formats export to clean Markdown

Key Features to Implement

1. Document Conversion

from docling import DocumentConverter

converter = DocumentConverter()
doc = converter.convert("path/to/file.pdf")
markdown = doc.export_to_markdown()

2. Hybrid Chunking for RAG

from docling.chunking import HybridChunker

chunker = HybridChunker()
chunks = chunker.chunk(doc)  # Semantic + token-aware chunking

3. Audio Transcription

from docling.pipeline import AsrPipeline

# Whisper Turbo for local audio transcription
# Supports 90+ languages with timestamps

Implementation Tasks

Core Integration

• Add Docling dependency to project
• Create document processing module using DocumentConverter
• Implement file format detection and routing
• Add error handling for unsupported formats

RAG Pipeline Enhancement

• Integrate HybridChunker for intelligent document splitting
• Configure token limits (typical: 512 tokens for embeddings)
• Preserve metadata (sections, headings, timestamps)
• Update vector database insertion to handle Docling chunks

Audio Processing (Optional)

• Add FFmpeg dependency for audio support
• Configure Whisper Turbo ASR pipeline
• Implement timestamp extraction for temporal referencing
• Support MP3, WAV, M4A, FLAC formats

Advanced Features (Future)

• Picture classification & description (IBM Granite Vision)
• Code syntax understanding for technical docs
• Advanced table structure recognition (TableFormer)
• Formula and diagram extraction

Benefits

• No Manual Parsers: Eliminate custom PDF/Word/Excel parsing logic
• Better RAG Performance: Semantic chunking improves retrieval accuracy
• Local Processing: Everything runs locally with Hugging Face models
• Fast: 30-second audio transcribed in ~10 seconds, complex PDFs in <30s

References

• Docling Documentation
• Docling GitHub
• Tutorial: Turn ANY File into LLM Knowledge
• Research notes in Slack #research channel

Technical Notes

• Replaces current document processing with unified API
• Compatible with existing pgvector/Pinecone/Qdrant implementations
• Token-aware chunking respects paragraphs, sections, tables
• Markdown output is ideal for LLM consumption