# Retrieval-Augmented Generation (RAG) > Retrieval-Augmented Generation (RAG) finds relevant pages in a company's own documents and gives them to an AI language model before it answers. It grounds a construction team's responses in its actual specs, proposals, and project history instead of the model's general knowledge. - Stages: Preconstruction, Estimating, Closeout - Concepts: Retrieval (RAG), Data - Published: 2026-08-28 - Canonical: https://buildr.com/library/retrieval-augmented-generation ## Why it matters in construction A language model knows what a Division 07 scope generally contains. It does not know what your Division 07 scope on the Meridian project contains, what your roofing sub excluded last time, or what your firm's standard exclusions are. Every useful precon answer depends on that firm-specific material. RAG gives AI tools access to that material without retraining. It supports questions about project documents, similar past estimates, and prior work for a client. When a vendor says its AI works with your data, it often means the product uses RAG. ## How it works 1. **Index.** The system splits each document into chunks of a few hundred words, usually along section boundaries, converts each chunk to an embedding (a numeric representation of its meaning), and stores it with metadata like project, document type, and page. 2. **Query.** When someone asks a question, the system embeds the question the same way and pulls the chunks whose meaning sits closest. Metadata filters narrow the search to the right project or document type. 3. **Assemble.** The top chunks, usually five to twenty, go into the prompt alongside the question and an instruction to answer only from the provided material. 4. **Generate.** The model writes an answer from those chunks and, if built well, cites which chunk each statement came from. 5. **Verify.** The person reads the answer with links to the source pages and checks anything that matters. An answer cannot use a page that retrieval never returns. Good chunking, clean metadata, and re-indexing when addenda arrive determine whether RAG has the right material to work from. ## Example in practice Suppose a GC has 11 years of estimates, sub proposals, and closeout documents in file storage. A precon manager pursuing a $27M cold storage facility wants to know what the firm carried for insulated metal panels on comparable jobs and which subs performed. With a RAG-based tool over the archive, she asks the question in plain language. The system retrieves the IMP sections from three past estimates, two sub proposals, and a closeout punch list. The answer: unit costs of $28 to $34 per square foot across 2022 to 2025, one sub with a warranty claim on panel joints, and another that finished two weeks early on the most similar project. Each figure links to the page it came from. Before, that research meant opening a dozen project folders and hoping the file names were accurate. ## Go deeper - [AI for Estimators: 5 Practical Uses that Work Today (Not Someday)](/blog/ai-for-construction-estimators.md) - [GPT Models for Contractors: A Plain-English Guide](/blog/gpt-models-for-contractors.md) - [How to Build a Subcontractor Database That Actually Gets Used Across Your Team](/blog/how-to-build-subcontractor-database.md) ## How Buildr applies this Kit answers questions about your pipeline, estimates, and project history by retrieving from your own Buildr data and citing where each answer came from. See [Buildr Platform](/platform). ## Related terms - [Embeddings and Semantic Search](/library/embeddings-and-semantic-search.md): Embeddings are numeric representations of text that place similar ideas near each other even when the wording differs. Semantic search uses them to find construction documents, subs, or past projects by meaning rather than exact keywords, so acoustical ceiling can match ACT and lay-in tile. - [Grounding and Citations](/library/grounding-and-citations.md): Grounding ties an AI answer to specific source material rather than the model's general knowledge. Citations show where each claim came from. In construction, that might mean pointing to Section 09 29 00, paragraph 3.4, for a level 5 finish requirement. - [Context Window](/library/context-window.md): A context window is the amount of text, measured in tokens, that an AI language model can consider in one request. It limits how much of a construction spec book or proposal set the model can see at once. - [Hallucination](/library/hallucination.md): A hallucination occurs when an AI language model states something confidently that is untrue or unsupported by the source. For example, it might invent a unit price or an exclusion that does not appear in a subcontractor's proposal. - [Natural Language Querying](/library/natural-language-querying.md): Natural language querying lets a construction team ask questions about business data in plain English, such as 'what is our weighted pipeline for Q4' or 'which superintendents are free in March,' and get answers from the CRM, backlog, and workforce plan without building reports or formulas. ## Referenced by - [AI Proposal Generation](/library/ai-proposal-generation.md): AI proposal generation uses language models to draft construction proposal content, such as approach narratives, past project descriptions, and team bios, from a general contractor's own history and the RFP requirements. The pursuit team then reviews and edits the draft. - [AI Subcontractor Matching](/library/ai-subcontractor-matching.md): AI subcontractor matching uses a general contractor's sub database and bid history to recommend which subcontractors to invite for each trade package on a construction project, based on trade, location, capacity, past performance, and qualification status. - [Fine-Tuning](/library/fine-tuning.md): Fine-tuning further trains an existing AI model on a specific set of examples so it learns a particular style, format, or task. A construction firm might use it to teach a model its conventions for classifying subcontractor scope items. - [Tokens](/library/tokens.md): Tokens are the small chunks of text that AI language models read and write, roughly three-quarters of a word each. They determine both the cost of a request and how much of a construction document fits in a model's context window. ## FAQ ### Is RAG the same as training the model on our data? No. Training changes the model itself and is slow and expensive. RAG leaves the model alone and gives it your documents at question time. It is often a better fit for construction document questions because those documents change regularly. ### What documents work well with RAG? Anything text-based that you would otherwise search: spec books, contracts, sub proposals, past estimates, meeting notes, closeout manuals, and CRM history. Drawings and heavily graphical content need different handling. ### Does RAG stop hallucination? It reduces it substantially because the model has the real source in front of it. It does not eliminate it. Retrieval can miss the right page, and the model can still misread what it retrieved, so citations remain important.