Building the World's Most Accurate AI-Powered Document Retrieval Platform: A 4-Year Journey
As the co-founder of il, I've spent 15 years building AI-powered applications, and now I'm excited to share our innovative approach to achieving industry-leading accuracy in Large Language Models (LLMs) for real-world document retrieval.
- 1. Ben is a co-founder of il, which builds AI-powered applications.
- 2. He has been working in this field for 15 years, starting at IBM research and then IBM Watson.
- 3. Il has developed the world's most accurate and scalable RAG (Retrieval Augmented Generation) platform using no-code tools and APIs.
- 4. Users can upload documents and receive accurate retrievals within minutes.
- 5. Il has been developing its solution for four years and was an early user of the GPT3 beta program.
- 6. RAG can have high error or hallucination rates, especially with complicated enterprise documents.
- 7. The source of these errors is usually not the language models or prompts, but RAG itself or the quality and relevance of retrieved content.
- 8. Content-related problems typically fall into three categories: bad text extraction, missing information, and non-extracted visual elements.
- 9. Most issues with RAG are related to content ingestion and can take hundreds of hours to solve.
- 10. Il has spent four years addressing these data engineering challenges and has built solutions into its ingestion pipeline.
- 11. As a result, users can build accurate RAG applications in just minutes.
- 12. Il's platform has achieved 98% accuracy against complicated real-world documents, outperforming some popular solutions by as much as 120%.
- 13. Il does not use vector databases, which they believe may not be the best technology for many RAG applications.
- 14. Instead, il creates what they call semantic objects and performs a multi-field search across their attributes.
- 15. Semantic objects include the original chunk text and auto-generated metadata that preserves information around the text.
- 16. Il rewrites the text into two ideal formats: one for search and another for completion.
- 17. For example, Air France has been using il's platform to build a chat GPT-like copilot for their call center agents, with a knowledge base of hundreds of thousands of documents.
- 18. Il's ingestion pipeline identifies images, tables, and text in the documents and processes them through dedicated multimodal processing pipelines.
- 19. Rewriting the query into a format compatible with semantic objects allows il to search the entire object: original text, auto-generated metadata, and search version of the text.
- 20. Il uses fine-tuned language models to rank results and improve accuracy.
- 21. In total, more than nine models are fine-tuned for ingestion and search in il's platform.
- 22. The end result is the world's most accurate RAG platform, enabling users to build enterprise-quality production-ready applications in minutes instead of months.
- 23. Il invites users to try their platform at il.la/xray.
- 24. Ben concluded his presentation by thanking the audience.
Source: AI Engineer via YouTube
❓ What do you think? What are the most significant challenges in building a high-accuracy Rag application, and how can we overcome these hurdles to achieve better performance? Feel free to share your thoughts in the comments!