RAG development: AI that answers from your own documents and data.
We build retrieval-augmented generation over your documents, knowledge base and live systems, deploy it on hosted, private or self-hosted models, and measure it against your real questions before anyone relies on it. One knowledge assistant from $8,000 fixed, live in 3 weeks.
What RAG development services should deliver.
Retrieval-augmented generation (RAG) means the model looks up the relevant passages in your own documents and data before it answers, and cites them. It is how you get answers about your policies, products, contracts or past work without training a model on them, and how the answers stay current when the documents change. The model is rarely the hard part. The work is in cleaning and chunking the corpus, retrieval that finds the right passage, permissions so people only see what they are allowed to, and an evaluation set that tells you whether the answers are right. Gigabit builds all of it, with a human path for the questions the system should not answer.
Three published builds use it. A Shopify Plus brand's support agent answers from its catalog, policies and live order data; before launch it scored 91.7% response accuracy and 99.1% policy compliance on 2,000 historical tickets, and it now resolves 64% of tickets on its own. A 140-person consulting firm's knowledge base covers 12,000 documents with access control by practice area and returns a sourced answer in 4 seconds. A dermatology clinic's patient assistant serves cited answers from clinician-reviewed SOPs and cut repetitive front-desk inquiries 62%. When data may not leave your environment, the same system runs on a model in your own cloud account or on your own servers.
What a RAG build includes.
Every build ships with an evaluation set, citations, access controls and monitoring, so you can see what it answers and how well.
Document and knowledge-base RAG
Answers over SharePoint, Google Drive, Confluence, Notion, PDFs and wikis, with structure-aware chunking, versioned sources and a link to the passage behind every answer.
Retrieval over live data
Answers that combine documents with records pulled at question time from your store, CRM, ERP or database, so order status and stock levels are never a day old.
Permission-aware retrieval
Each user retrieves only what their role, team or client authorization allows, enforced at query time, with PII redacted before text reaches the model.
Private and on-premise LLM deployment
Hosted models in your own cloud account and region, or open-weight models served on your own GPUs or servers, behind your network and your logging.
Evaluation and monitoring
A test set built from your real questions, scored before launch and after every change to prompts, models or the index, with alerts when quality drops.
Fine-tuning, when it earns its place
A smaller model trained for one narrow, high-volume job such as classification or a fixed output format, compared on your eval set against retrieval and prompting first.
From your documents to answers you can check.
Questions and corpus
Collect the questions people actually ask, audit the sources that answer them, and agree what the system must refuse or hand to a person.
Prove retrieval
Build the eval set first, then tune chunking, embeddings, hybrid search and re-ranking until the right passage comes back for each question.
Integrate and deploy
Answers with citations inside Slack, the help desk, your product or a web widget, on the model and hosting your data rules allow.
Operate
Re-index as documents change, review failed and escalated questions, and re-run the eval set on every model or prompt change.
Published prices, mapped to the job.
Every price here is the same one on the pricing page. Work with no published price gets a fixed price in writing before it starts.
Figures in the notes are what past clients paid, from the published case studies; they show the range, not a rate card. Work with no natural end, such as ongoing development, is priced per engineer per month with no lock-in.
When to hire us for this, and when not to.
- Staff or customers ask the same questions, and the answers sit in documents nobody can search.
- Answers have to cite their source, because someone will check them.
- Contracts, regulation or a customer's security review say the data may not leave your cloud or your building.
- Someone can say what a correct answer looks like for the questions that matter.
- Your help desk, intranet or document tool already has a built-in AI search that covers your sources. Switch it on and test it first.
- The documents contradict each other and nobody owns them. Fix the source of truth first; retrieval will only repeat the contradictions.
- You want a model fine-tuned on your documents so it knows them. For facts that change, retrieval is cheaper to run and easier to keep correct.
Retrieval in production, with the numbers.
A support agent grounded in catalog, policy and order data, a consulting firm's knowledge base over 12,000 documents, and a clinic assistant serving cited answers.
A confidence-gated AI support agent now resolves 64% of tickets autonomously, cutting support costs 62% — $217,200 saved in year one.
A phased AI transformation recovered 1,200 consultant-hours per quarter and $480K/year for a 140-person consulting firm — 129% Year 1 ROI.
A RAG patient assistant with clinical guardrails cut repetitive front-desk inquiries 62% and time-to-answer 68% in 8 weeks.
These are past engagements at the prices those clients paid. For build price, monthly cost and payback side by side, see what ten client projects cost.
RAG development, answered
What do RAG development services cost?
One knowledge assistant over your documents, with up to three tool integrations, is the AI Quick Win Sprint at $8,000 fixed over 3 weeks. A production agent that answers and acts starts at $8,000 as a flat fee; our published support agent, grounded in a 420-SKU catalog, policies and live order data, cost $42,000 to build. Private and on-premise deployments get a fixed price in writing after scoping, because hardware and hosting vary. The $25,000 AI Transformation Sprint finds where retrieval pays off first, and monitoring after launch runs on Managed AI Operations at $3,000–$20,000 a month.
Should we use RAG or fine-tune a model?
Start with retrieval for anything factual. RAG answers from the current version of your documents, cites them and changes when they change; a fine-tuned model learns from a snapshot and cannot show where an answer came from. For the published support agent, fine-tuning was rejected because it would have baked in the inconsistencies in historical agent answers. Fine-tuning fits a narrow, high-volume task where a smaller model must follow one format or label set reliably. Provider options also change: on 24 October 2026 OpenAI's documentation says it is winding down its fine-tuning platform, which no longer accepts new users. We compare options on your eval set before recommending one.
Can we run an LLM privately or on premises?
Yes. There are three common setups: a hosted model from a major provider run inside your own cloud account and region; an open-weight model served on GPUs in your cloud account; or an open-weight model on your own servers. Open-weight models are served through software such as vLLM, which exposes an OpenAI-compatible API, so the application does not change if you switch models later. Each model has its own licence to check: the Llama 4 Community License, for example, requires a separate licence from Meta above 700 million monthly active users and a visible "Built with Llama" notice (checked on 24 October 2026). We run your eval set on the private model and a hosted one, so you see any quality gap before you decide.
How accurate is a RAG system?
As accurate as its retrieval and its sources, and you should see the number before launch. The published support agent was tested against 2,000 historical tickets with known resolutions and scored 94.2% intent classification accuracy, 91.7% response accuracy and 99.1% policy compliance; it only replies on its own when retrieval relevance is above 0.80 and intent confidence above 0.85, and routes everything else to a person with a draft. Your system gets its own eval set built from your questions, re-run after every change.
Which vector database do you use?
The one that fits the data you already run. If you are on Postgres, pgvector adds vector similarity search with exact and approximate nearest-neighbour indexes to the database you already back up and secure; the consulting firm's knowledge base runs on it. A managed vector service such as Pinecone, which the support agent uses, suits teams that want no index to operate. We usually pair vector search with keyword search, because product codes and names are matched better by exact terms.
How do you keep a RAG system secure?
Permissions are enforced at retrieval time, so a user cannot get an answer built from a document they could not open. Retrieved text is treated as untrusted, because a document or web page can carry instructions aimed at the model, and the model's tools get least-privilege access with logging of every call. For a formal test before launch, the AI Security Assessment red-teams the model, the retrieval layer and the tools, from $12,000 fixed.
Often scoped alongside.
AI integration
Models and agents built into the software and workflows you already run.
AI integrationGigabit Agents
Production AI agents for a flat fee, from $8,000.
Gigabit AgentsData engineering
Pipelines and clean, documented tables that retrieval and AI features can rely on.
Data engineeringWhat is MCP?
How the Model Context Protocol connects agents to your systems once.
What is MCP?From the field notes
Have documents your team or customers should be able to ask?
Tell us where the documents live, who asks the questions and whether the data may leave your environment. We reply with what we would build first and the fixed price.