AI that answers from your own data, and shows its sources.
Assistants and agents built on your documents and tools, measured against a test set we agree first.
An AI assistant answering a question from your own documents and citing its sources.
Overview
RAG (retrieval-augmented generation) lets an AI model look things up in your documents before it answers, so replies come from your data and cite where they came from. We build the pipeline that cleans and indexes your content, the assistant or agent that uses it, and the checks that tell you when it gets something wrong. The model is swappable: Claude, OpenAI or open-source models you host yourself. Sometimes the right answer is no RAG at all: a few hundred pages that rarely change can go straight into the prompt.
A good fit if
- Your answers live in documents, tickets and systems that change every week.
- Each answer has to show where it came from.
- Different staff may see different documents, and the assistant has to respect that.
- You want the assistant to take actions in your tools, not only answer.
Not the right start if
- Your content is a few hundred pages that rarely change. It can sit in the prompt. We'll tell you on the first call.
- Nobody owns the source documents. Sort out ownership first. An assistant only repeats what it's given.
- You already have an AI demo that people don't trust yet. Start with AI Pilot to Production
Retrieval
Watch a question find its answer
Every answer starts with a search. Switch parts of the search off and see what the assistant would miss.
Can we get a refund on an annual plan after 30 days?
Meaning search
Keyword search
- Cancellation FAQ: Monthly plans end at the next billing date.
- Pricing: Annual plans cost less per month than monthly plans.
- MissedRefund policy 4.2: Annual plans can be refunded within 60 days of purchase.
Rerank
Answer
Refund policy 4.2 Cancellation FAQ No source found
Answered from Refund policy 4.2, with the passage linked.
Answered from Refund policy 4.2, with the passage linked. Meaning search missed the exact phrase, so the refund rule never reached the model. The assistant says it can't find a rule, which is the safe failure. The refund rule was found but ranked third, so the answer came from the cancellation FAQ, which is the wrong one. A reranker reads the candidates and puts the best one first. Neither safeguard is on. The refund rule is missed and the assistant says it can't find one.Every case
- Both on
- Answered from Refund policy 4.2, with the passage linked.
- Keyword search off
- Meaning search missed the exact phrase, so the refund rule never reached the model. The assistant says it can't find a rule, which is the safe failure.
- Reranker off
- The refund rule was found but ranked third, so the answer came from the cancellation FAQ, which is the wrong one. A reranker reads the candidates and puts the best one first.
- Both off
- Neither safeguard is on. The refund rule is missed and the assistant says it can't find one.
5.7% to 1.9% Giving each passage its surrounding context, then adding keyword search and a reranker, cut failed retrievals from 5.7% to 1.9% in Anthropic's tests. Anthropic, 2024
Process
How it runs
You keep
- 01A test set of real questions and a refusal list.
- 02A source and access map.
- 03The ingestion pipeline and its chunking rules.
- 04The search index and a retrieval score on your test set.
- 05Citation and refusal rules.
- 06A versioned test suite and a baseline scorecard.
- 07A tool list and approval rules.
- 08A trace dashboard and a runbook.
-
01 of 08
List the questions.
We collect the questions your team really gets, and the ones the assistant must refuse.
You keep A test set of real questions and a refusal list.
-
02 of 08
Map sources and access.
Which systems hold the answers, who may see what, and how often each source changes.
You keep A source and access map.
-
03 of 08
Clean and split.
A parser for each format, tables and PDFs included, with every chunk labeled with where it came from.
You keep The ingestion pipeline and its chunking rules.
-
04 of 08
Search two ways.
Keyword and meaning search together, then a reranker that puts the best passage first.
You keep The search index and a retrieval score on your test set.
-
05 of 08
Answer with sources.
Every claim links to its passage, and the assistant says so when no source exists.
You keep Citation and refusal rules.
-
06 of 08
Score it.
Search and answers are scored separately, on every change.
You keep A versioned test suite and a baseline scorecard.
-
07 of 08
Set what it may do.
For agents: the tools it can call, what needs a person's approval, and when it must stop.
You keep A tool list and approval rules.
-
08 of 08
Run it.
Traces, user feedback and an alert when a source stops syncing.
You keep A trace dashboard and a runbook.
Risks
Most wrong answers start with the search
-
The search misses the passage.
The model can only answer from what the search hands it.
Prevented in step 4 -
A confident invention.
In April 2025 a software company's support bot made up a login policy, and customers cancelled over it.
Prevented in step 5 -
An agent with too much power.
In July 2025 a coding agent deleted a live production database during a code freeze.
Prevented in step 7
5.7% to 1.9%
Giving each passage its surrounding context, then adding keyword search and a reranker, cut failed retrievals from 5.7% to 1.9% in Anthropic's tests.
Deliverables
What you keep
-
RAG pipeline over your data
Ingestion, chunking and search tuned to your documents, with every answer linked to its source.
-
Agents that use your tools
Assistants that call your APIs and systems, with limits on what they may do alone.
-
Evaluation suite
A test set and pass mark we agree with you, rerun on every change.
-
Cost and latency monitoring
Dashboards for spend per answer and response time, with alerts.
-
Model-agnostic build
The model sits behind one interface, so you can swap providers without a rewrite.
Done means
- Every factual answer links to its source.
- Questions outside your sources get a refusal, not a guess.
- Staff only get answers from documents they're allowed to see, and that's tested.
- Anything that changes data waits for a person's approval.
- The test suite passes the mark we agreed before the build.
Sample
Test scorecard
| Questions in the test set | 120 |
|---|---|
| Answered from the right source | 108 |
| Refused correctly | 9 |
| Wrong or unsupported | 3 |
| Right passage in the top 5 results | 96% |
| Pass mark agreed before the build | 90% from the right source |
FAQ
Questions about AI Agents & RAG
Still deciding?
Ask the people who would build it. We reply within one business day.
Not always. If your content is a few hundred pages and rarely changes, it can go straight into the prompt, which is simpler and cheaper to run. RAG earns its place when the content is large, changes often, or must show its sources.
It can, so we design for it. Answers must cite a source, the assistant refuses when none exists, and people are told they're talking to AI, which the EU AI Act has required since 2 August 2026.
RAG for knowledge, fine-tuning for behavior. If the model lacks your facts, retrieval fixes that. If it knows them but answers in the wrong format or tone, fine-tuning can help, and we try better prompts before either.
We agree a test set and a pass mark with you before we build. If the pilot misses it, you decide whether we adjust the approach or stop, and you keep everything we built up to that point.
Yes. Your data stays in accounts you control, and we use model providers' business tiers that don't train on your data. Only the people on your project can access it.
Yes. We connect it through your APIs or through MCP (Model Context Protocol, an open standard for giving AI models tools), with the same permissions your staff have. Anything that changes data waits for approval.
Contact
Have an idea or a stuck project? Let's talk.
We reply within one business day. If we're not the right team, we'll say so on the first call.
On the first call
- The questions you want answered, and where the answers live today.
- Who may see which documents.
- Whether you need RAG at all.