Your enterprise data is not in the model. RAG puts it there.

RAG (Retrieval-Augmented Generation) retrieves relevant documents from a knowledge base at query time and passes them to the language model as context before generating an answer. The model answers the question based on the retrieved documents, not solely from its training data. Fine-tuning adjusts the model's weights by training it on your data, updating what the model 'knows.' The practical differences are important. RAG keeps your data in a retrieval system you control: documents are indexed, not baked into model weights. When your documents change, you update the index. When a document is removed or superseded, it is removed from the retrieval system and the model stops citing it. Fine-tuning bakes knowledge into the model: updating it requires retraining, and there is no clear source citation. For most enterprise use cases, customer support, policy lookup, product information, legal review assistance, RAG is the right approach because the knowledge changes frequently and source attribution matters. Fine-tuning is more appropriate for changing the model's reasoning style, output format, or task-specific behaviour.

RAG works with any content that can be indexed: PDFs, Word documents, PowerPoint presentations, web pages, Confluence and Notion pages, Zendesk or Intercom knowledge base articles, plain text files, and structured databases. The ingestion pipeline extracts content from the source format, chunks it into segments appropriate for retrieval, generates embeddings, and stores them in a vector database. For structured data (databases, spreadsheets, CSVs), we use hybrid retrieval approaches: semantic search over unstructured content combined with structured query generation for database records. The retrieval quality depends on document quality and structure. Well-organised, clearly written documents retrieve better than dense, poorly structured ones. We assess your document library during scoping and identify any document quality or organisation issues that will affect retrieval before we commit to retrieval quality targets.

Retrieval quality measures how often the retrieval system finds the right documents for a given query. A RAG system can generate fluent, confident-sounding answers from retrieved documents and still be wrong if it retrieved the wrong documents. Retrieval quality has two dimensions: recall (does the system retrieve the documents that contain the answer?) and precision (does the system avoid retrieving irrelevant documents that confuse the answer?). We evaluate retrieval quality using a test set of questions and expected source documents, measuring recall and precision at different retrieval depths. We iterate on chunking strategy, embedding model, and retrieval configuration until retrieval quality meets a defined threshold for your use case before building the answer generation layer on top of it. Retrieval quality is the foundation. Everything else depends on it.

Document-level access controls are a first-class design requirement in enterprise RAG. The retrieval system must only surface documents the querying user has permission to read. We implement access control in the retrieval layer using metadata filtering: each document in the vector index is tagged with its access group metadata, and at query time the retrieval query is filtered to only return documents the current user's permissions allow. For RAG systems connected to existing document repositories (SharePoint, Confluence, Google Drive), we use the source system's permission model: a document the user cannot read in SharePoint is not indexed for retrieval by that user. Access control design is defined during scoping and tested before deployment. A RAG system that leaks confidential documents to users who should not see them is a more serious problem than one that retrieves the wrong document.

We price land-and-expand. A focused first use case (for example, a customer support knowledge base with one document set) starts around $25,000. That is the smallest credible slice: one workflow, one source, one retrieval quality target, live and proving value. From there a multi-use-case enterprise platform with hybrid retrieval, reranking, and role-based document access grows toward $60,000 and up. We scope the work, calculate the cost, and lock it in writing before development starts. No sliding invoice.

Yes. We sign NDAs before any technical discovery or document sharing. Most clients across the US, UK, Europe, Canada, and the UAE operate under NDAs from the first call. We have handled confidential document sets for legal firms, healthcare operators, and financial services teams. The NDA is signed before we see any content.