How Speculative Decoding Cuts OCR Hallucinations by 90%

Video Description

This talk explains why document parsing is still hard and why vision-first methods outperform OCR for complex layouts, merged cells, charts, signatures, and messy long-tail formats. Reducto shows how their pipeline combines computer vision models with a vision-tuned LLM and speculative decoding to reduce hallucinations, preserve reading order, and create reliable structured output for Elasticsearch. Developers learn how to build more accurate document-driven RAG systems using CV-first parsing. Chapters: 00:00 Why document parsing fails 00:30 OCR limits 01:00 Vision-first approach 01:30 Combining CV and LLMs 02:00 Speculative decoding 03:00 Structured output 04:00 Elasticsearch integration 05:00 Final example Additional resources: - Introducing Elastic’s Agent Builder: https://www.elastic.co/search-labs/blog/elastic-ai-agent-builder-context-engineering-introduction - Expose Agent Builder tools to any AI agent: https://www.elastic.co/search-labs/blog/elastic-mcp-server-agent-builder-tools Start the 14-day trial for free! No credit card required: https://cloud.elastic.co/registration Subscribe to Elastic’s Community YT channel: https://www.youtube.com/c/OfficialElasticCommunity Connect with us on social media: LinkedIn: https://www.linkedin.com/company/elastic-co X: https://twitter.com/elastic Facebook: https://www.facebook.com/elastic.co About Elastic Elastic, the Search AI Company, enables everyone to find the answers they need in real time, using all their data, at scale. Elastic’s solutions for search, observability, and security are built on the Elastic Search AI Platform — the development platform used by thousands of companies, including more than 50% of the Fortune 500. #ElasticSearch #ElasticObservability #ElasticSecurity