We are thrilled to announce a major milestone: Progress Agentic RAG is officially ranked as the #1 top result on the Snowflake MADQA Leaderboard, taking the top spot across both Agentic RAG and conventional RAG evaluation benchmarks.
This milestone represents the culmination of an ambitious vision. Exactly one year ago, Nuclia was acquired by Progress. Since that day, our team has been working relentlessly to transform enterprise information retrieval by embedding deep agentic capabilities into every single layer of the RAG pipeline from ingestion and content enhancement to hybrid retrieval and federated integration. This is the Agentic Knowledge Layer at work: not one trick that boosted a single benchmark, but a governed foundation across ingestion, retrieval, and evaluation, now validated by a public, third-party leaderboard.
What Is the MADQA Leaderboard, and Why Does It Matter?
MADQA (Multimodal Agentic Document Question Answering) is a benchmark built by Snowflake AI Research with the University of Oxford, UNC-Chapel Hill, CVC, and Hugging Face. It tests systems against 2,250 human-authored questions grounded in 800 real enterprise documents (financial reports, forms, technical diagrams) many of which require reading visual structures like tables, checkboxes, and stamped dates, not just extracted text.
What makes MADQA a harder test than most RAG benchmarks is that it doesn't just grade the final answer. It logs the entire search trajectory — every query, every page view — and compares how a system navigates a document collection against how a human researcher does. It's also a closed-world: systems can't lean on outside training knowledge, they have to find and cite real evidence in the provided files. Snowflake's own research found that even the best agents needed up to 9 rounds of search to match the accuracy humans reached in 1–2, meaning most systems get there by brute force, not efficient reasoning.
That's why a #1 finish here means something beyond a leaderboard number: it's independent, third-party evidence that an answer is both accurate and properly grounded which is precisely the standard enterprise buyers, compliance teams, and now AI agents themselves are starting to hold retrieval systems to.
Here is an inside look at how we built the top-performing RAG solution on the market and what this means for enterprise AI.
Agentic Capabilities Across Every Layer of RAG
Traditional RAG systems treat data ingestion, processing, and search as isolated, rigid steps. We took a radically different approach by deploying intelligent agentic mechanisms across every phase of the data lifecycle.
Ingestion: Intelligent Layout & Multi-Format Extraction
Unstructured data comes in complex shapes including spreadsheets, presentation slides, multi-column PDFs, embedded images, and intricate data tables.
- Layout Awareness: Progress Agentic RAG implements dynamic extraction strategies that automatically analyze documents to identify and apply the optimal layout model for each specific file type.
- Complex Data Handling: Whether parsing multi-sheet workbooks or multi-modal diagrams within slides, the system isolates, structure-maps, and preserves the fidelity of complex content automatically.
Content Enhancement: AI-Ready Data Preparation
Garbage in, garbage out. High-performing retrieval depends entirely on data quality:
- Smart Splitting: Rather than relying on arbitrary chunk sizes, our splitters analyze context boundaries to preserve semantic coherence.
- Proprietary State-of-the-Art Embeddings: We leverage custom-built embeddings fine-tuned for both general-purpose enterprise contexts and domain-specific scenarios.
- Data Augmentation & Knowledge Graphs: Using advanced data labeling alongside semantic and LLM-driven graph extraction, we automatically construct knowledge graphs to map relationships between concepts, entities, and metadata.
Retrieval: The Agentic Loop in Action
Retrieval isn't just vector matching; it's an active, iterative reasoning process.
- Ground Truth + Vector Search: We combine deterministic, extractable structured metadata (for absolute ground truth) with high-dimensional vector search.
- Intent & Filter Intelligence: Every query undergoes intent detection, dynamic metadata filtering, and automated query rephrasing.
- Graph Traversing: The retrieval engine actively traverses semantic knowledge graphs to pull multi-hop context that traditional vector search completely misses.
Taking It Further: The Federated Agentic Framework
Creating the best AI-ready data for enterprise knowledge bases was step one. Next, we solved the challenge of connecting these capabilities to the broader enterprise ecosystem.
We recently published our Federated Agentic Framework, enabling next-level enterprise automation:
- Federated Connectivity: Seamlessly bridge internal enterprise data stores with external enterprise tools and databases.
- Protocol Support: Built-in support for Model Context Protocol (MCP) and Agent-to-Agent (A2A) interaction patterns, enabling autonomous AI agents to converse, delegate, and execute complex workflows safely.
- Federated Security: End-to-end security and granular access controls ensure that data boundaries and permissions are strictly enforced, no matter how many agents or systems join the flow.
Enterprise Proven: Reliable, Fast, and Flexible Deployment
We have always known that our platform delivered the fastest, most reliable, and easiest integration history in the industry. But enterprise AI requires deployment flexibility without compromising control:
- Multi-Cloud Support: Native, optimized deployments across AWS, GCP, and Azure (both public and private clouds).
- Air-Gapped Environments: Complete offline support for strict security, defense, or highly regulated industry requirements.
The Verdict is In
We always knew we had built the most reliable and performant RAG architecture. Now, the Snowflake MADQA Leaderboard offers public, verifiable proof. It also lands as the market validates the whole category: 67% of production LLM deployments now use some form of retrieval augmentation, up from 31% in 2024 (McKinsey, 2026 State of AI in Enterprise), and the rise of autonomous agents is the single biggest driver of that jump. Agents need somewhere trustworthy to retrieve from and now there's public, third-party proof of which layer does it best.
By combining intelligent multi-format extraction, graph-enhanced semantic indexing, agentic retrieval loops, and a federated agent ecosystem, Progress Agentic RAG delivers unmatched accuracy and verifiability.
Check out the rankings yourself on the Snowflake MADQA Leaderboard on Hugging Face!
Learn more about Progress Agentic RAG by booking a demo, and take advantage of our new QuickStart Pro offer of 100k tokens/month for the first three months of your usage before the offer expires.