Answers from your own data, within each person's access
AI assistants and enterprise search that only show what each person may see
We build AI chatbots and enterprise search over your own documents and data: answers that cite their source, retrieval that respects each user's permissions, and an evaluation set that shows when answers get worse. On Databricks that means AI Search indexes governed by Unity Catalog, Genie for questions about tables, and apps that query as the signed-in user.
- Who it is for
- IT and data leaders asked for a company assistant that can't leak
- Bench
- 115+ engineers, averaging 15 years each
What we bring with us
Our own accelerators, running in your workspace.
Evaluation and quality
The evaluation harness. A fixed set of real questions scored on every change, so a new model or a re-chunked index is shown to help before anyone relies on it.
Agent development framework
For the assistant that has to do something as well as answer, such as raise a ticket or draft a reply. Harness runs it as an agent with approvals and runs that survive a restart.
All of it is free and Apache-2.0 licensed, and it runs inside your own workspace under your Unity Catalog. We deploy and run it during an engagement, or you take the source and run it yourself. See the whole family.
A chatbot that answers from a folder of documents takes an afternoon now. The hard part arrives the first time it answers with a paragraph from a document the person asking was never allowed to open. Permissions have to travel with the data into the index and back out through the app, and that's where most of these projects stall.
What we work out first
- Which questions people actually ask, taken from the help desk queue, a shared inbox or a week of search logs
- Where the answers live: documents in SharePoint, tables in the lakehouse, records in the ERP or the CRM
- How each source holds its permissions, and whether they survive being copied into an index
- Which questions are about documents, which are about numbers and which need both, because each is built differently
- What a wrong answer costs, which decides how much the assistant may say without a source behind it
What we build
- Document search on Databricks AI Search, formerly Vector Search: Delta Sync indexes that stay current with their source tables, and hybrid keyword and vector retrieval so part numbers and codes still match
- Questions about numbers answered by Genie over governed tables, with the table and the query behind each answer on screen
- Apps that use Databricks Apps user authorization, so the signed-in person's row filters and column masks decide what comes back
- Answers that cite the passage or the table they came from, so checking one means reading the source
- An evaluation set of real questions with known answers, rerun whenever the model, the prompt or the index changes
- The assistant where people already work: a Databricks App, a web application, or inside a product you run
Clients we have done this for
Manufacturing
A governed Databricks lakehouse that joins Business Central with the scale house, the line historian, the lab and market prices, with nine leadership dashboards and a Genie space, plus Dext expenses posting into Business Central without re-keying.
AI systems · Our own product
Fabric
Fabric is a context store, memory and agent runtime built on Databricks, and it is the clearest evidence of what we build: across our clientele, work that took a team of ten now takes three.
Where this goes next
Agentic AI development
If the assistant has to take actions, not only answer
Genie Accuracy
If the questions are about numbers and Genie gets them wrong
Data integration
If the answers live in an ERP or CRM nobody has landed yet
Data science and AI
Context and memory underneath an assistant that runs for months
A Genie app that won a Databricks challenge
Ivan Vydrin's Prove It shows the query before the result
FAQ
AI assistants and enterprise search: common questions
What does an AI chatbot development company actually build?
The parts around the model. Retrieval that finds the right passage, permissions that decide what each person may see, citations so an answer can be checked, an evaluation set that shows when it gets worse, and the interface people use every day. The model is bought; the rest is engineering, and that's the part we do.
What is AI enterprise search?
Search across a company's documents and data that answers a question in a sentence, with its sources, instead of returning ten links. It combines keyword search, which finds exact terms like a part number, with vector search, which finds passages that say the same thing in different words. Databricks AI Search does both and merges the results.
How do you stop a chatbot showing documents someone shouldn't see?
Permissions are applied when content is retrieved, never filtered out of an answer afterwards. On Databricks, AI Search indexes are governed by Unity Catalog, and an app using user authorization queries as the signed-in person, so their row filters and column masks apply. Documents copied in from elsewhere carry their access lists as metadata, and every query filters on them.
Should we buy an enterprise search product instead?
Often, yes. If your documents live in a handful of SaaS tools and an off-the-shelf product connects to all of them, buying is faster and cheaper.
Building pays off when the answers depend on data in your lakehouse, ERP or CRM, when the permissions are unusual, or when the assistant has to sit inside a product you already run. We'll tell you which one you're looking at in the first conversation.
Can it answer questions about numbers, not just documents?
Yes, by a different route. Numbers come from Genie or a query over governed tables, with the query shown, rather than from a model summarising a PDF of last quarter's report. In the plant reporting build we've published, a Genie space answers plain-English questions about line downtime with the chart and the table behind it.
How do you know the assistant is any good?
A set of real questions with known answers, scored every time the model, the prompt or the index changes. Without it, quality is whoever spoke last in the review meeting, and nobody can safely move to a newer model.
The rest of our AI work
AI Readiness Assessment
Leaders with AI on the roadmap and no clear view of what is in the way
AI Tools Assessment
Finance and platform leaders holding a renewal they cannot justify
Agentic AI development
Engineering leaders whose agent pilot cannot get past review
AI workflow automation
Operations and engineering leaders with a process that half-completes
AI enablement
Engineering leaders whose team needs to own this, not outsource it
AI transformation
Executives with a mandate, a budget, and fifteen pilots