I will build an ai knowledge agent for your company documents with sourcea


About this gig
WHAT YOU GET
An AI knowledge agent that answers questions about YOUR documents employee handbook,
policies, price lists, quality manual, technical documentation. Every sentence of every
answer carries a footnote with the document name and page number, and you can open the
cited passage with one click.
WHY NOT JUST UPLOAD A PDF TO A CHATBOT
Three reasons my clients come to me after trying that:
1. No sources. You cannot check whether the answer is real. Mine cites document and page.
2. Your documents leave the building. In my setup, reading and indexing run locally on your
own machine only the question and the retrieved paragraphs go to the language model.
3. It guesses. A general chatbot will invent a plausible number rather than admit a gap.
My agent has three separate locks against this:
· If nothing in your documents is close enough to the question, the language model is
never asked in the first place.
· The model is required to cite a source for every sentence; where there is no source,
the prescribed answer is "this is not in the documents".
· A second, independent pass checks every statement against the passage it cites and
marks anything that goes beyond it.
Get to know Lukas.W
n8n Automation Developer for German Businesses, GDPR Compliant
- FromGermany
- Member sinceMay 2026
- Avg. response time1 hour
Languages
English, German
FAQ
Does my documents leave the company ?
Reading, splitting and the vector index run locally on your own machine. Per question, only the question itself and the retrieved paragraphs (usually 6 to 12) are sent to the language model. A fully offline setup with a local language model is possible – answer quality drops with the smaller model,
What does it cost to run?
You use your own API key. Roughly 1 to 3 cents per question; 500 questions a month is about 5 to 15 euros.
What happens when a document changes?
You drop the new file in the folder and run one command to rebuild the index. I can also set that up to run automatically.
Can it answer questions that need two documents at once?
Yes – the search runs across all documents, and the answer cites each source separately. If two documents contradict each other, the agent says so instead of picking one.
What if it gets something wrong?
That is what the test report is for: we measure it before you accept. Within the revision rounds I tune the splitting and the retrieval until the agreed questions pass.

