Use casesKnowledge assistants
An assistant over your own documents that answers with the page, or admits it has none.
Policies, procedures, product sheets, ten years of email. The answer is in there somewhere, and someone spends an afternoon finding it. We build knowledge assistants that answer only from your documents, cite the passage every time, and say plainly when the documents do not say.
What gets in the way
A new hire asks how an exception is handled. The answer is in a procedure from 2019, a chat thread and an email nobody can find. The expert who knows is in a meeting.
Point a model at the document store and it answers everything, including the questions the documents do not cover. It cannot tell you which version it read, and it cannot tell you it is guessing.
For a team that lives on getting things right, an assistant is only worth having if every answer can be checked in one click and every silence is honest.
What we build
Retrieval finds the passages. The model writes only from them. Code checks that it did.
- Documents are indexed with their version, owner and date, so an answer from a superseded policy is caught at retrieval, not by a reader.
- The model is given the retrieved passages and nothing else, and every sentence it writes has to point at one of them. A sentence with no passage behind it is cut.
- A separate check, with no model in it, confirms the cited passage actually contains the claim, the same way we verified every citation in Hand & Stone’s database without asking a model.
- Access follows your permissions. A person who cannot open a document cannot get its contents through the assistant.
- Unanswered questions are logged. They are the list of documents you need to write.
How we know it works
The exam is a set of real questions with known answers, including questions the documents cannot answer.
- Before launch, your experts write a few hundred questions and the right answers. The assistant is graded on them, and the misses are reviewed one by one.
- A share of the questions are traps: topics the documents do not cover, questions that depend on a superseded version, questions outside a user’s permissions. Each must end in an honest “not in the documents”.
- Citation verification without a model named the right source on 1,230 of 1,239 checkable records in the Hand & Stone work. The same check runs on every answer here.
- The document review demo on this site shows the rule at small scale: the answer, the sentence it came from, the gaps and the conflicts.
Built from this work
We have not shipped an internal knowledge assistant under a client’s name yet. The retrieval, the citation rule and the verification without a model are in the work below.
1,193 schools in the finished database
572 new schools found beyond Hand & Stone’s original list
68 of 201 top-tier schools that were not on the original list
Read the story
Where it fits
Insurance
Insurance, from the question a customer asks to the systems that answer it.
Financial services
AI for financial services, built to pass an audit.
Franchising
AI for franchise brands, researched at a scale no team does by hand.
Questions
Which document stores?
SharePoint, Google Drive, Confluence, a file share, a wiki. Anything we can read with your permissions. Each source keeps its version and its owner.
What about documents that contradict each other?
The assistant shows both passages and says they disagree, with the dates, instead of picking one. That is usually the moment a policy gets cleaned up.
Does our data train a model?
No. The documents are retrieved at question time and never used to train anything. The knowledge stays yours.
How do we keep it current?
The index refreshes on a schedule from the source of truth. Delete a document and its answers stop. The unanswered-questions log tells you what to write next.
Bring us the project that matters most.
Tell us what you’re building and what’s in the way. You’ll hear back within 24 hours.