Services › AI Agents › Knowledge assistant
AI Agents
An AI knowledge assistant that answers from your company documents — with sources
Your team asks a question in plain language and gets an answer drawn from your own documents, with a link to the file it came from. No more digging through drives, wikis and old email threads.
What is an AI knowledge assistant?
An AI knowledge assistant is an internal tool that answers employees' questions using the company's own documents. Instead of searching folders by keyword, people ask in plain language — "what's our refund policy for annual plans?" — and get a short answer with links to the exact sources. It only answers from documents it is allowed to read, says when it can't find something, and stays current as the documents change.
Do you need a knowledge assistant?
- "The answer exists somewhere, but nobody remembers which folder."
- "New hires ask the same questions for months, and senior people lose hours answering them."
- "Our procedures live in Google Drive, Notion and a few PDFs, and they don't agree with each other."
- "Sales and support keep asking the product team the same technical questions."
- "We tried searching our wiki. Keyword search doesn't understand what people mean."
What is RAG, and why not fine-tune a model?
Retrieval-augmented generation (RAG) means the assistant first finds the relevant passages in your documents, then asks the language model to answer using only those passages. The documents are split into small sections and stored in a vector database, which can find passages by meaning rather than exact words. Because the answer is built from retrieved text, the assistant can show you exactly where each answer came from.
Fine-tuning means retraining a model on your data. For a knowledge assistant it's usually the wrong tool: it's more expensive, has to be repeated whenever documents change, can't reliably cite sources, and still makes things up when it's unsure. RAG keeps your knowledge in your documents, where you can update it in minutes.
| RAG | Fine-tuning | |
|---|---|---|
| Uses your latest documents | Yes — re-sync and it's current | Only after retraining |
| Shows the source of an answer | Yes | No |
| Respects who may see what | Can, at retrieval time | Hard — knowledge is baked into the model |
| Setup cost | Lower | Higher, and repeated |
| When it makes sense | Answering from documents | Changing a model's style or a narrow, stable task |
What do we build?
- Connectors to your document sources, with scheduled syncing so new and edited files are picked up.
- A search index in a vector database, organised so answers come from the right team's documents.
- The assistant itself, with instructions to answer only from retrieved sources and to say "I don't know" otherwise.
- Citations on every answer, linking to the original file or page.
- An interface where your team already works — for example a simple web app, Slack or Microsoft Teams.
- An admin view showing which questions go unanswered, so you know which documents to write next.
How does a knowledge assistant answer a question?
- Question. An employee asks in their own words.
- Retrieve. The system finds the passages whose meaning matches the question best.
- Filter. Passages from documents the person isn't allowed to see are dropped.
- Answer. The model writes a short answer using only the remaining passages.
- Cite. Links to the source documents appear under the answer, so anyone can check it.
Which document sources can it use?
We connect the sources your team already uses. For example:
Each source has its own API and permission model, so we confirm during the audit exactly what can be synced and how often.
What about permissions, freshness and privacy?
Access permissions
An assistant must not become a back door to documents someone shouldn't see. The simplest safe setup separates documents by team or access level. Where a source's API exposes per-file permissions, the assistant can check them for each question. We agree the approach before building, because it shapes the whole design.
Keeping content fresh
Documents are re-synced on a schedule, and deleted files are removed from the index. Answers are only as good as the documents behind them, so the admin view highlights questions with no good source.
Private and self-hosted options
If documents must not leave your infrastructure, the workflows can run on self-hosted n8n with an open-source model hosted privately. It costs more to run than a commercial API, and we'll explain the trade-off for your case.
How long does it take and what does it cost?
A first version connected to one or two sources typically takes 4–6 weeks. Setup starts at $3,000, and monthly care starts at $300 per month. Model and hosting usage is paid to the providers and depends on how many documents and questions you have.
Monthly care covers keeping syncs running, adding new sources, tuning answers based on the questions people actually ask, and fixing anything that breaks when a source tool changes. The build is paid in milestones, and the launch includes one month of free bug fixing.
When is a knowledge assistant not a fit?
Good fit when…
- Your knowledge is written down, even if it's scattered
- The same questions come up again and again
- People lose time searching or interrupting colleagues
- You can say who should see which documents
Not a fit when…
- Most knowledge lives only in people's heads
- You have a handful of documents a search box handles fine
- Documents are badly out of date and nobody owns them
- Answers must be legally binding without human review
Frequently asked questions
What is a RAG chatbot for company documents?
A RAG chatbot answers questions by first retrieving the relevant passages from your company's documents and then asking a language model to answer using only those passages. Because answers are built from your own text, it can show the source of each answer and stays current when documents are updated.
Is RAG better than fine-tuning a model on our documents?
For answering questions from documents, usually yes. RAG is cheaper to set up, picks up changes as soon as documents are re-synced, can cite its sources and can respect access rights. Fine-tuning is better suited to changing a model's style or teaching a narrow, stable task.
Can employees see documents they shouldn't through the assistant?
Not if it's designed properly. Documents can be separated by team or access level, and where a source exposes per-file permissions the assistant can check them for every question. The approach to permissions is agreed before the build, because it shapes the whole system.
Which tools can an AI knowledge assistant connect to?
Common sources include Google Drive, Notion, Confluence, SharePoint, PDFs and internal wikis. The exact connections depend on each tool's API and your plan with that vendor, and are confirmed during the audit before any work starts.
Can the knowledge assistant run privately on our own servers?
Yes. The workflows can run on self-hosted n8n and the model can be an open-source one hosted privately, so documents never leave infrastructure you control. Running costs are higher than with a commercial API, so we compare both options for your case.
Is your team answering the same questions every week?
Tell us where your documents live. We'll tell you what an assistant could answer, and what it would take to build.
Book a Free Call →