How we help / LLM and RAG development
LLM and RAG development services for your own documents and data
We build LLM applications that answer from your documents, records and systems rather than the open internet, with the evaluation, permissions and cost controls that let people rely on them.
Most of the work in an LLM feature is not the model. It is getting the right content to it, checking what comes back, and keeping it safe and affordable once real users arrive. Building a whole AI product? See generative AI development. Want the model to take actions in your systems? See AI agent development.
What you get
Custom LLM and RAG development covers these parts. Each project weighs them differently, depending on the content, the users and the cost of a wrong answer.
- The right model
- Hosted models from the major providers or open-weight models in your own cloud, chosen by running your own questions against two or three candidates and comparing quality, cost per answer and where the data may go.
- Retrieval over your content
- Ingestion, parsing and chunking of your documents, wikis and application data, with hybrid search, reranking and an index that stays current when a source changes. See how to build a RAG chatbot.
- Prompts and tools
- Instructions that keep answers to the retrieved sources, cite them and say plainly when the answer is not there, plus safe, predefined tools for questions that need a database query rather than a document search.
- Fine-tuning, when it pays
- Considered only after prompts and retrieval have been tried and evaluation shows a gap they cannot close, such as a strict output format or a high-volume task a smaller tuned model could do for less.
- Evaluation and guardrails
- A test set built from real questions, checks on retrieval and on answers, and input and output checks that stop off-topic or badly formed responses before they reach a user.
- Security and permissions
- Access rules applied at retrieval time, so people only get answers from content they may already read, with personal data kept out where it is not needed and retrieved text never allowed to override instructions.
- Deployment and monitoring
- Running in your cloud or ours, with every question logged alongside what was retrieved and what was answered, so a bad answer can be traced and fixed.
- Cost control
- Tokens, latency and cost per answer tracked by feature, fewer and better chunks in each prompt, caching where answers repeat, and a cheaper model wherever the test set shows it is good enough.
How it goes
Five steps, in this order. You see answers on your own questions early, not only at the end.
- Collect real questions
Gather the questions people actually ask, with the correct answer and its source for each. This becomes the test set that every later decision is checked against.
- Prepare the sources
List what the system may answer from, who owns it and how often it changes, remove superseded versions, and check by eye that parsing keeps tables and structure intact.
- Prove retrieval and pick the model
Build a working prototype, measure whether the right passages are found, and compare candidate models on the test set before committing to one.
- Build the complete system
Add permissions, citations, guardrails, the interface people will use and the ingestion pipeline that keeps the index current, with the evaluation suite running on every change.
- Launch small and watch it
Start with one team or a share of users, review failures every week, track cost per answer, and widen the rollout once the numbers hold.
Why fine-tuning is rarely the first step
Fine-tuning changes how a model behaves, not what it knows today. Facts that change belong in retrieval, where they can be updated, removed and restricted the same day.
- Start with prompts and retrieval
- Most answer-quality problems turn out to be retrieval problems: the right passage was never found, or an old version was. Fixing those is cheaper than training anything.
- Measure the gap first
- Fine-tuning is worth discussing when the evaluation suite shows a consistent failure that better prompts, examples and retrieval have not fixed.
- Count the running cost
- A tuned model needs training data, retraining when the base model changes and its own evaluation. We weigh that against what it saves per answer before recommending it.
Questions people ask first
What is the difference between RAG and fine-tuning?
RAG retrieves relevant passages from your content for each question and asks the model to answer from them, so the knowledge stays in an index you control and can update at any time. Fine-tuning trains a model further on examples, which changes its style and behavior but is a poor way to keep facts current. For answering from business documents and data, RAG is almost always the right starting point.
Should we use a hosted model or an open-weight model?
It depends on answer quality on your own questions, cost per answer, and where your data is allowed to go. Hosted models are quick to start with and strong on quality. Open-weight models run in your own cloud account, which suits strict data residency rules, but you take on hosting and scaling. We test the candidates on your test set and design the system so the model can be changed later without rebuilding the rest.
How much does LLM or RAG development cost?
There is a build cost and a running cost. The build cost depends mostly on the content and the integrations: how many sources, how messy the documents are, whether permissions have to be synced from other systems, and where the answers appear. The running cost is mainly model usage, hosting the index, and the time people spend reviewing answers. We do not quote a fixed price before we have seen the content. Our guides on building a RAG chatbot and AI agent development cost explain how to estimate both.
Can people see content they should not have access to?
Not if permissions are applied at retrieval. We store access rules with each piece of content and filter the search by the user’s identity and groups before anything reaches the model, then test with users in different roles. Hiding content after the model has seen it is too late.
How do you know the answers are right?
By measuring them. We evaluate retrieval and answers separately, using a test set built from your real questions, checks that every claim is supported by the retrieved sources, and cases the system should refuse. The suite runs on every change, including model updates on the provider’s side. Our guides to LLM evaluation metrics and evaluating LLM applications in production describe the approach.
Do you build from scratch, or adapt something that exists?
Either. If you need a chatbot that answers from your documents, URLs and application data, ChatData in our AI Studio may already fit as it stands or with changes for your sources and systems. When nothing existing fits, we build the LLM application ourselves, starting from one use case.
Where we have done it
Guides from our blog
Have documents or data people keep asking about?
Tell us what the content is, who asks the questions and what a wrong answer would cost. We will reply with what we would build first.