The problem
The answer exists in a policy, a ticket, a PDF or a chat thread. Finding it takes longer than asking the person who has been there for years.
New joiners ask the same questions for weeks. The answers are in someone’s head or in a document nobody bookmarked.
A generic chatbot gives confident answers with no source, which is worse than no answer in a support or compliance context.
There is no test set, so nobody knows whether the assistant is improving or quietly getting worse after a document update.
What you get
Outcomes
- Fewer repeated questions and faster onboarding.
- Answers with citations, so the reader can check the source.
- Permissions that follow the documents, not a single shared index.
- A measurable quality baseline on real questions.
- A feedback loop and a runbook for adding or removing sources.
Deliverables
- A source inventory: documents, tickets, policies and the owner of each.
- An ingestion and chunking strategy for each source type.
- Retrieval with citations and source links.
- Permission handling for sensitive collections.
- An evaluation set of 30 to 80 real questions with a pass/fail rubric.
- A feedback mechanism for wrong or missing answers.
- Deployment in your environment, with logs and a runbook.
- A handover session and a plan for keeping the sources current.
Evidence
Fit and boundaries
This is a good fit when
- 50 or more documents, tickets or policy pages.
- Repeated questions that already have written answers.
- An owner for the source content.
- A permission model, or a willingness to define one.
This is not a fit when
- No written sources. The assistant would invent the knowledge.
- A public-facing chatbot with no moderation or escalation path.
- A single shared index for documents with different access levels.
- No willingness to review the evaluation failures before launch.
How it works
- 01
Collect the real questions
The questions people already ask in support, onboarding and internal chat. A test set built from invented questions measures the wrong thing.
- 02
Fix the sources first
Stale or contradictory documents produce confident wrong answers. The source audit happens before the model choice.
- 03
Build retrieval and citations
The assistant must show where an answer came from. Citations are part of the product, not a debug feature.
- 04
Evaluate and launch
The evaluation set is scored before launch, and the failures are reviewed. The harness stays with your team.
Pricing and timing
Fixed-price engagement
€4,000 to €8,000 3 to 4 weeks · EUR, VAT not applicable at present
Fixed price, three to four weeks, including the evaluation set and deployment in your environment.
- Written scope and acceptance criteria before work starts.
- 40 to 50 percent deposit reserves the calendar slot.
- Milestone invoices and 14-day payment terms.
- A checkpoint where either side can stop with paid work settled.
Questions
Can it answer from our own documents?
Yes. The assistant retrieves from the sources you provide and cites them. It is not trained on your documents by default; retrieval keeps the answers traceable and the sources updatable.
How do you stop hallucinations?
No system can promise zero wrong answers. Citations, refusal behaviour, a retrieval audit and an evaluation set make the failures visible and reduce the rate. The harness is how you keep it honest after launch.
Can different teams see different documents?
Yes, with a permission model at retrieval time. A single shared vector index is the wrong default for sensitive collections.
Does it integrate with Slack, Teams or our helpdesk?
It can. The assistant is delivered as an API first, then connected to the interface your team already uses. That keeps the integration replaceable.
Page last updated: October 2026