RAG copilots for company knowledge
Short answer
SwankyForge builds retrieval-augmented generation (RAG) copilots: assistants that answer staff or customer questions from your own documents, policies and systems, and cite the passage each answer came from. Every release is checked against a test set of real questions before it goes live. A typical project takes 4–8 weeks and costs €8k–€18k.
Can a customer return an opened item after 20 days?
Yes, within 30 days if the item is undamaged and in its packaging. Opened hygiene products are excluded.
- [1] Returns policy, section 4.2
- [2] Hygiene products list, updated in March
2 sources · answer checked against the release test set
Sounds familiar?
- Staff interrupt colleagues instead of reading the policy.
- Support answers the same questions every day.
- A chatbot pilot answered confidently and wrongly.
- Knowledge is spread across wikis, PDFs and tickets.
- Nobody can tell whether the assistant is getting better or worse.
What we build
Internal knowledge assistant
Policies, procedures and product documentation, answered with a link to the source.
Support answers
Draft replies for agents, or direct answers to customers, from your help centre.
Search across sources
PDFs, wikis, tickets and shared drives in one place, with their permissions.
Lookups in your systems
Questions about an order, a contract or an account answered from live data.
Release tests
A suite of real questions that every change has to pass.
Access rules
People only see answers built from documents they are allowed to open.
How a project runs
- 1
Collect real questions
A few hundred questions people actually ask, with the answers an expert would accept. This becomes the test set.
- 2
Connect the sources
Documents, wikis and tickets are indexed with their permissions, so nobody sees a document they could not open.
- 3
Retrieve, then answer
The system finds the relevant passages first, then writes an answer that quotes and links them. No source, no answer.
- 4
Measure before release
Each change is scored on the test set: correct, partly correct, wrong or declined. A release that scores worse does not ship.
- 5
Run and improve
Logging, feedback buttons and a weekly review of failed questions feed the next round.
What you get
- Assistant in the channel your team already uses: web, chat or helpdesk
- Test set of real questions with expert answers
- Evaluation report for every release
- Source and permission connectors
- Code, prompts and runbooks in your repository
Typical price
€8k–€18k
Typical timeline
4–8 weeks
From the published rate card. The final price is fixed once the scope is agreed.
Estimate your projectQuestions
RAG or fine-tuning: which one do we need?
For answering from company knowledge, start with RAG: it uses current documents, shows its sources and updates without retraining. Fine-tuning helps with tone, format or a narrow task, and is usually a later step, if it is needed at all.
Will our data be used to train someone else's model?
No. We use model providers under terms that exclude training on your data, or run open models inside your cloud when that is the requirement.
How do you stop it from making things up?
It answers only from retrieved passages and declines when nothing relevant is found. The test set measures how often it is wrong, so the risk is a number rather than a guess.
Which language model do you use?
Whichever scores best on your test set at an acceptable cost. The system is built so the model can be swapped without a rewrite.
What does a RAG assistant cost to run?
Model usage is paid per question. We estimate it from your expected question volume before the project starts, and the dashboard shows the actual figure afterwards.
Tell us about the task.
The first call is free. If this system is not the right answer for your numbers, we say so.