W hen you hear the words “filling out a form,” what comes to mind? Tedium? Repetitive data entry? The soul-crushing task of copying information from one document and pasting it into a series of boxes? You’re not alone. For decades, forms have been the symbol of manual, uninspired work.
But what if we told you that the humble form is actually a complex challenge for AI? What if we said that a perfectly completed form is not just a filled-in document, but a demonstration of deep, contextual understanding?
At Ontos, we believe that to solve the biggest problems in knowledge work, you must first master the fundamentals. And for AI, the ultimate test of its ability to extract, reason, and integrate knowledge lies within the simple, deceptive structure of a form. This is why we created SnapFill, and it’s why we started here.
The Form: A Disguised Quest for Knowledge
On the surface, a form is just a set of empty fields. But for an AI, a well-designed form is a structured query into a complex body of knowledge. A grant application isn’t just asking for a budget; it’s a comprehensive inquiry into a research project’s core components, viability, and potential impact. A new-hire onboarding packet is a structured request for a person’s complete professional and personal identity.
Simple automation tools fail here because they operate on a “copy-paste” level. They can match keywords, but they can’t _understand_ context. This is where the real challenge begins, and it’s what separates basic automation from true intelligence.
Why Forms Are the Ultimate AI Gauntlet
To intelligently complete a complex form, an AI must do more than just find words in a document. It must navigate a sophisticated web of information, much like a human expert does.
1.Going Beyond Keywords: Conceptual Alignment
It’s rarely a one-to-one match. A source document might talk about the “project budget,” while the form asks for “total funding required.” A simple keyword search for “total funding required” would come up empty. A truly intelligent system needs to understand that “budget” and “funding” are conceptually the same in this context. It needs to align concepts, not just words.
2.Understanding the Blueprint: Hierarchical Structures
Knowledge isn’t flat; it’s layered. An AI must grasp these parent-child relationships. It needs to know that a “Lead Researcher” is a type of “Team Member,” or that a “monthly expense report” is a component of the “annual budget.” Without this vertical understanding of knowledge, it can’t correctly categorize and place information. It might list a project lead’s salary under general team expenses, missing the crucial hierarchical distinction.
3.Connecting the Dots: Uncovering Hidden Relationships
This is where most AI systems falter and where the deepest understanding is required. The most critical information is often found in the relationships _between_ the data points:
- Sequence: When a form asks for “Project Milestones,” the AI must read a project plan and discern the correct chronological order of tasks. Task A must come before Task B.
- Causality: If a field asks for the “Project’s Novelty,” the AI can’t just find the word “novelty.” It must understand that _Technology A_ leads to _Performance Improvement B_, and then articulate that cause-and-effect relationship.
- Comparison: For a “Market Competition Analysis” section, the AI needs to sift through multiple reports, identify competing products based on similar features, and summarize their relevant strengths and weaknesses in relation to your own.
This intricate dance of vertical, horizontal, and conceptual understanding is what makes form-filling a grand challenge. An AI that can conquer this has proven it doesn’t just process words — it comprehends knowledge.
Our Answer: SnapFill, The Intelligent Drafting Assistant
Because of this complexity, we chose forms as the first and most critical application for our core technology at Ontos. We built SnapFill not as another “autofill” gimmick, but as an intelligent drafting assistant powered by a deep understanding of your knowledge.
This brings us to the story of why we started this journey. During my PhD and postdoctoral research, I, like many academics and professionals, drowned in paperwork. Grant proposals, administrative reports, progress updates — I felt like I was filling out the same information in slightly different boxes hundreds, if not thousands, of times a year. My co-founder shared this frustration from his own experiences. We saw how much human potential was being wasted on this high-volume, low-value work.
We knew there had to be a better way. We envisioned an AI that could act as a true assistant — one that could read all our project documents, understand them deeply, and then draft a perfect, context-aware form for us to simply review and submit. This personal pain point became our mission. We decided to become the “saviors” for researchers, and for all professionals, drowning in forms.
The Ontos Engine: How We Taught AI to Truly Understand with Tree-RAG
So, how do we do it? The magic behind SnapFill lies in our foundational technology, a novel approach to Retrieval-Augmented Generation (RAG) that we call Tree-RAG.
For years, the world of AI has been trying to solve the problem of feeding external knowledge to Large Language Models (LLMs). The common approach, often called “Naive RAG,” is like tearing a book into thousands of tiny, disconnected sticky notes. You chop up documents into small paragraphs, toss them into a database, and hope the AI can find the right sticky note when you ask a question. The result? Fragmented, out-of-context, and often inaccurate answers.
A more advanced method, “Graph RAG,” tried to solve this by creating a complex “map” of how concepts relate to each other. This was a step forward, but it was incredibly expensive, slow to update, and required significant manual effort to design. Every time new knowledge arrived, you had to redraw the entire map.
We knew neither of these approaches could handle the dynamic, ever-changing nature of real-world knowledge. Our solution is Tree-RAG.
Imagine instead of a pile of sticky notes or a rigid map, your knowledge is organized like a book with a dynamic table of contents. This is what Tree-RAG does.
- Builds the Tree: It starts by analyzing your documents and understanding their inherent structure — chapters, sections, subsections. This preserves the original context from the very beginning.
- Fuses New Knowledge: When new information arrives — from an email, a meeting transcript, or a chat message — Tree-RAG doesn’t just throw it into the pile. It intelligently analyzes the fragment, understands its meaning, and seamlessly inserts it into the correct branch of the “book.” A note about a “database configuration error” is automatically filed under “Databases → Common Issues → Configuration.”
- Retrieves with Precision: When it’s time to fill a form, SnapFill uses a hybrid search that looks at both the content (what you’re asking for) and the structure (where it lives in the book). This dual-engine approach ensures the answer is not only accurate but also contextually complete.
In our tests on real-world IT operations data, Tree-RAG outperformed other methods dramatically. It boosted retrieval accuracy by 13%, updated knowledge 2–3 times faster than Graph RAG, and reduced token consumption by up to 90%. Most importantly, it was the only system that consistently improved as more knowledge was added, never getting noisier or less accurate.
Are you ready to experience the efficiency of true understanding?
Join the waitlist for SnapFill today and be the first to witness the revolution in intelligent document processing. Let’s end the tyranny of manual data entry, together.
Originally published on Medium.
Blog 