Notebook LM for construction? Meet the new AI Agent from BinderLess

Artificial intelligence is increasingly entering construction sites. However, the real change doesn't start with robots or drones, but with... documents. This is where project managers, inspectors, and engineers lose hours of work every day - hours that can be reclaimed.

BinderLess introduces a solution that can change this - the world's first AI Agent fully integrated with a construction documentation platform (CDE). With it, teams can simply talk to their documentation and get answers in seconds that previously required tedious searching through hundreds of files.

Before we look at how our AI Agent works, it's worth understanding the basic concepts behind its capabilities. In this article, we explain the fundamentals of artificial intelligence that form its foundation - such as LLM, RAG, visual grounding and self-reflection loop - and show how they translate into real benefits in daily work on the construction site and in the design office.

Artificial Intelligence (AI) is a broad set of technologies that allow computers to perform tasks requiring human-like intelligence. This can include language understanding, i.e., reading comprehension, image recognition, data analysis, or decision-making.

An AI Agent is a specific type of AI that not only processes information but acts with a specific purpose, responds to user queries, and makes decisions in the context of a given task. For the Construction AI Agent, this means:

• Understanding natural language questions - you can ask, for example, "What is the heat transfer coefficient for the window in zone C?" and the Agent will understand the meaning of the question.
• Data search and analysis - The AI Agent from BinderLess searches through project documentation in the CDE platform, combines different sources, and generates a response based on actual data, not assumptions.
• Contextual interaction - The Agent remembers the project context, user role, and previous queries to ensure responses are precise and useful.
• Information supplementation and generation - In later stages, the Agent will be able to create reports, suggest problem solutions, or automatically fill in missing data in documents.

In short: AI in the form of an Agent not only answers questions but becomes an active partner in the work process, helping to save time, avoid errors, and efficiently manage documentation.

LLM is the foundation of AI Agents and serves as its "language brain". Thanks to it, the Agent can converse in natural language, interpret user intentions, and generate readable responses, which are then supported by actual data from project documents.

LLM, or Large Language Model, is a type of artificial intelligence that can understand and generate natural language. These are machine learning models trained on enormous text datasets (books, articles, websites), through which they learn:

1. Statistical relationships between words and sentences

The model knows which words and phrases usually occur together and in what context. Here's a simpler example from construction, showing statistical relationships in LLM: "On the construction site, it's best to keep a helmet on ..." - The model looks at statistical relationships in the training data. Most often after such a beginning, the following appear:

• "head" - 12%
• "desk" - 8%
• "shelf" - 4%...
• "lynx" - 0%

Statistically, in texts about safety, helmets are most often kept on the head, so the model will most likely complete the sentence with the word "head". In short: LLM chooses the answer that most often appears in a similar context in the training data, even if there are other possible variants.

Fun fact

* This is a simplification, as the model also takes into account temperature, which is a parameter controlling the randomness (i.e., "creativity") of how subsequent words are selected. If the temperature is 0, the model will always choose the word with the highest probability, there's no randomness here. If temperature = 1, the model sometimes chooses a different word — still probable, but not always the same, making its responses more natural and varied. The higher the temperature, the more creative and unpredictable the model becomes, but at the same time, the risk increases that the answer will be less sensible or less factual.

2. Creating coherent responses

It can generate sentences that are logically correct and sensible in a given context. This means that the model can not only choose individual words but generates entire, coherent sentences or paragraphs in response to questions. For example:

• "How to prepare the substrate for a concrete screed?"
• "The substrate should be cleaned of dust and loose elements, the surface should be leveled and moistened with water to ensure proper adhesion of the concrete."

This answer is logically correct and sensible in the context of the question, even if the model hasn't seen this exact project. LLM combines fragments of knowledge from various documents and patterns, creating a comprehensive, understandable description of the action.

3. Understanding user intent

Thanks to patterns in the text, the model can assess what the user is asking about, even if the question is not formulated perfectly precisely. For example:

• To the question: "What margins for windows?", which is informal and abbreviated, the model "guesses" based on training data that the word "margins" in the context of "windows" usually refers to allowable installation tolerances and based on this knowledge generates the answer:
• "The allowable tolerances for window installation in masonry walls are usually ±5 mm, depending on the standard and type of construction."

In short: LLM can "filter" imprecise or abbreviated questions and interpret them correctly, providing a sensible and useful answer.

In practice, LLM operates in two stages:

• Input – the model receives text (e.g., a question, document fragment).
• Output – the model generates a response, predicting subsequent words based on what it learned during training.

Examples of LLMs include Chat GPT from Open AI, Gemini or Notebook LM from Google, and Claude from Anthropic.

RAG in Construction – Intelligent Responses Based on Documentation

Despite the fact that LLM is trained on large datasets, it doesn't "know" anything about specific project documents – it operates based on patterns from training data, meaning it lacks context for a specific construction. For the Agent to be useful in construction, it's necessary to combine the LLM system with the Retrieval-Augmented Generation (RAG) system.

What is RAG?

RAG is a technique that combines large language models (LLM) with the ability to access external information sources, e.g., project documentation. Thanks to this, the AI Agent doesn't rely solely on training patterns but refers to specific, real data, providing precise answers.

In practice, RAG works like this:

1. User query goes to the LLM, which interprets the question's intent and formulates a "contextual query" to the database or project documents.
2. Retrieval – the system searches the documentation in the CDE platform, selecting fragments most related to the question.
3. Augmented Generation – LLM combines its linguistic knowledge with the found fragments and generates a full, coherent response, linked to actual documents.

Thanks to RAG, the AI Agent:

• doesn't "hallucinate" answers – it always refers to the source,
• provides precise and up-to-date information, even if the documentation is scattered,
• allows the user to ask questions in natural language and receive answers consistent with the project and standards.

In practice, this means that:

• LLM interprets and processes language – it understands questions in natural form, even imprecise or abbreviated ones.
• RAG provides precision and project context – the Agent refers to actual document fragments in the CDE.

AI Agent Connected to the CDE Platform - why is it Important?

In construction, language understanding alone is not enough. What matters here is data – drawings, documentation, specifications, dozens of file versions. That's why BinderLess based its AI Agent on the RAG architecture.

Now that we're aware of what LLM is and why without proper RAG (retrieval-augmented generation) the answers would be imprecise, it's worth understanding what connecting such an AI Agent with a CDE (Common Data Environment) platform provides.

Unlike a regular GPT chat, which can answer general knowledge questions, the AI Agent in CDE uses current, verified project documentation, allowing it to provide precise, contextual answers tailored to a specific building and its data. When you ask, for example, "What is the heat transfer coefficient for the window in section B2?", the Agent analyzes the documentation, finds the appropriate fragment, cites it, and provides the source. It doesn't guess. It really knows where the information comes from.

In practice, this means that artificial intelligence in construction projects can become an active participant in the information flow – one that understands documents, connects data, and provides answers within seconds.

Self-reflection Loop and Visual Grounding – how AI Learns to Understand Projects

Connecting the AI Agent with the CDE platform opens up entirely new possibilities, going beyond a regular GPT chat. Thanks to integration with full project documentation and visual construction data, the Agent not only answers questions but can understand the project context, indicate connections between documents, and support real-time decision-making.

Understanding the basic concepts and operation of the AI Agent, we can now focus on several advanced features that make its use truly practical in everyday work: from visual reference to documentation, through self-correcting conclusions, to intelligent information linking and proactive suggestions in the project.

Additional Concepts Related to the Functioning of the AI Agent

Visual grounding – AI's ability to link text with an image or document fragment. In practice, this means that the AI Agent doesn't just say "the coefficient is in the table on page 42", but shows the exact fragment of the document where this information is located. This allows the user to immediately verify the answer and be sure it comes from the right source. In the context of construction, this is a huge change. Instead of searching through hundreds of PDF pages, the Agent visually indicates the place it's talking about. This speeds up work, reduces errors, and builds trust in the data.

Example: you ask "Where is the electrical installation conduit in the basement?" and immediately get a highlighted fragment of the floor plan.

Self-reflection loop – The biggest difference between traditional search and an intelligent AI Agent is that the Agent doesn't stop at the first answer found. Instead, it analyzes its own conclusions, checks their accuracy, and – if necessary – searches further and corrects.

Example: it detects an inconsistency in the specification and suggests a correction in the response.

Smart linking / dynamic citation – AI's ability to create hyperlinks and references to exact fragments of documentation. In practice, this means that each answer can lead directly to the source of information, allowing the user to immediately verify the data and gain full context. In construction, this eliminates time-consuming searching through PDFs and tables, providing quick access to the right drawing, specification, or document.

Example: The Agent responds "The insulation coefficient is in the table on page 42" and immediately leads to the highlighted fragment of the document.

Interactive suggestion / memory – the AI's ability to remember previous interactions and proactively suggest next steps or potential issues. In practice, this means that the Agent not only answers questions but also suggests what to check next, reminds about related issues, and maintains a consistent project context over time. In construction, this helps avoid recurring errors and facilitates coordination between teams.

Example: The Agent remembers that the electrical installation in the basement was previously checked and now suggests verifying the wiring route in adjacent rooms.

Contextual chaining – The AI Agent is not limited to a single document but can connect information from multiple sources, creating a coherent conclusion. This ensures that responses consider the entire project context, not just fragmentary data.

Example: It analyzes technical specifications and assembly instructions in various documents to determine that changing the material in one section of the project requires adjusting procedures in other parts of the documentation.

Multi-modal reasoning – The AI Agent can simultaneously analyze different types of data – text, tables, or charts – and combine them into a logical conclusion. This makes its answers more comprehensive and accurate than when analyzing a single type of data.

Example: It checks text documents and parameter tables to determine whether a selected component meets safety requirements and design standards.

The combination of these mechanisms makes AI more than just a search tool. It becomes an intelligent collaborator that can analyze, improve, and explain its logic in a way that is understandable to humans.

Data Security and Access Control – Trust at the Center of AI

Understanding how LLMs and AI Agents work, the next question that comes to mind is the issue of data, how it is used, and who has access to it on the platform.

Secure Use of Data in LLM

Unlike GPT chat, which learns from user-input data in a general model, our AI Agent works exclusively on documentation provided by the user and does not use it to train any global models. All data remains within the CDE platform, stored on secure servers in the EU, and only authorized project users have access to it.

Roles and Permissions - Project Restrictions

In construction, each project involves multiple roles and data sources. The project manager needs quick insight into work progress, the inspector checks compliance with standards, and the contractor looks for execution details. Project documentation - including drawings, specifications, and technical documents - is often confidential and scattered across various formats and locations.

The AI Agent doesn't operate in isolation but utilizes the entire documentation environment, aggregating information and understanding the project context. This means all files are available to the Agent, but the question remains: how to ensure that access to information is appropriately controlled and aligned with project roles?

Data Security in BinderLess

The BinderLess AI Agent has been designed to understand the context of each user and project. As a result:

• The Agent only has access to the data that the user on whose behalf the Agent is working has access to, meaning that an inspector only has access to documents relevant to their inspection.
• The Agent restricts access to other projects and organizations, which means that a manager cannot see confidential data from other projects and organizations.

In practice, this means that when a user asks a question, the AI automatically:

1. Recognizes which project and section the user is working in.
2. Takes into account their role and permissions.
3. Searches documents in the CDE, combining answers with the appropriate project context.
4. Displays the result along with a document excerpt for immediate answer verification.

This makes the AI Agent a secure, contextual assistant that integrates knowledge from the entire project and delivers it exactly where it's needed. Without worrying that the response will contain information from another project or that the user will see data they shouldn't have access to.

BinderLess AI Agent - your Documentation Assistant

In the construction industry, documentation is the heart of the project - it contains decisions, changes, and agreements, but in large projects, its volume makes finding needed information time-consuming. The BinderLess AI Agent solves this problem comprehensively. Through integration with the CDE platform, it utilizes full documentation, analyzes it using LLM and RAG, combines data from various sources (contextual chaining), can interpret text and images (multi-modal reasoning), and pinpoints exactly where in the documents the sought information is located (visual grounding and smart linking).

The Agent checks its answers, corrects them when necessary (self-reflection loop), and remembers the context of previous questions (interactive suggestion / memory). All of this operates within the permissions assigned to users, and data remains secure on EU servers, never entering global models or the internet.

The effect is simple but significant: less time spent searching for information, faster decision-making, reduced errors, and smooth access to project knowledge - exactly when you need it. The Agent doesn't replace people but streamlines the use of knowledge they already possess.

Converse with your Documentation

With the help of the BinderLess AI Agent, you can talk to your documentation as if it were a team member. Instead of opening dozens of PDF files and clicking through pages, you simply ask:

"What are the allowable tolerances for steel construction in zone C?"

The Agent understands intentions, analyzes documents, and draws conclusions. This allows it to deliver the most relevant information, not random results. It's not just a new way of working with documentation - it's a change in the entire flow of information on the construction site.

The Future of AI Agent in Construction - What else Awaits Us?

The development of the BinderLess AI Agent progresses in stages, from initial tests to features that already make it a full-fledged member of the project team. Each stage allows us to test new capabilities, gather user experiences, and gradually increase the scope of intelligent functions, making the Agent increasingly useful in daily work.

• Stage 0 - Test version. The beginning of the Agent's journey included a chatbot with a user guide. Tests allowed us to check how AI supports onboarding, quick access to knowledge, and user self-service without interacting with project documents.
• Stage 1 - Conversation with the Agent (September 2025). An Agent with basic document reading and searching functions within a specific project. Users can test intelligent search and interactive chat, receiving quick, precise answers.
• Stage 2 - Data creation capability (Fall/Winter 2025). In this phase, the Agent will start generating defect and complaint reports, as well as automatically creating or modifying project data. If necessary, the Agent asks the user for missing details, allowing for quick and accurate documentation completion.
• Stage 3 - Future development directions (2026+). Plans include features tailored for field work: voice reporting, voice assistant, access to external knowledge sources, playback of Agent responses in audio form, and support for field activities.

The development of the AI Agent is an ongoing process. We observe how users utilize its features, look for new use scenarios, and gather feedback to better respond to the needs of teams on site. Based on this, we will introduce new functions, expand our AI Agent's capabilities, and broaden its role in daily work, making it an even more indispensable assistant and decision support in every project.