Overview
Generic chatbots know nothing about your policies, so they fill gaps with guesses. We build an assistant that searches the documents you hand over, for example manuals, policies, help articles and contracts, stored in a vector index, then tuned using test questions you sign off. Every reply links to the passage that supports it, and when your documents do not cover a topic, the assistant says so plainly. We measure accuracy and refusals with real numbers instead of relying on a demo.
What you receive
- Import of PDFs, word documents, web pages and internal wiki content
- A vector index whose chunk sizes and embeddings suit the way your documents are written
- An answering service that links each reply to its source passage
- An embeddable website chat window, or a bot inside your team chat tool
- A 50-question test set, scored before tuning and again afterwards
- Safeguards that handle off-topic queries and block prompt injection
- A monthly running-cost estimate based on your expected query volume
How it runs
-
Step - 01
Gather
Your documents are collected, duplicates removed and the content cleaned up.
-
Step - 02
Build
We implement document import, search and answer generation end to end.
-
Step - 03
Evaluate
We score replies using your approved questions, then tune search quality.
-
Step - 04
Launch
A pilot group uses it first, then everyone once accuracy stays steady.
Timeline: 22 working days, counted from the day we receive what we need from you.
Is it a fit?
A good match for
- Support teams giving identical policy answers over and over
- Organisations with internal documents that are impossible to search
- Products where new users must wade through a lengthy manual
What we need from you
- Your documents, with the latest version of each one clearly marked
- Fifty genuine user questions along with their correct answers
- Your chosen language model provider, agreed before build starts
Why it helps
- Replies are tied to your documents and cite a passage anyone can verify
- Your most common questions no longer turn into support tickets
- A measured accuracy score to share, not just a demo impression
Questions
Grounding answers in your documents and refusing when unsure cuts invented answers sharply, and we measure how often it still happens. No such system can guarantee it never happens, so we share the measured rate.
No. We use providers whose terms exclude training on your data, or a model hosted on your own infrastructure where internal policy demands it.
That depends on how many questions are asked and which model is chosen. We prepare a cost estimate for your expected usage before you sign off on the project.