Product / Ask and Chat
Answers with sources, or an honest “not found”.
In short
troveGEN answers questions from your documents and shows where each claim came from. Totals are calculated from tables rather than guessed, every answer is scored for how well its sources support it, and questions the documents cannot answer are declined.
Trust
Checked citations
Every citation marker is validated against a real passage. A marker that points nowhere is removed.
Support score
Each answer is scored for how well its sources support it, and arithmetic in an answer is re-checked exactly.
Declines instead of guessing
When the documents do not cover a question, the answer says so. Related facts are still offered when they exist.
Exact numbers
Sums, averages, counts, minimums and maximums over tables are computed, then explained in words, with the rows used shown.
Using it
Chat for the whole team
Colleagues you invite chat with the projects you choose. Threads are private to each person; administrators see usage counts, never content.
Choose the answer model
Use any OpenAI-compatible model, including self-hosted ones, or Azure. Pick a model per message in Chat.
Follow-up questions
Threads remember earlier turns, so “what about the second one?” works.
Streaming
Answers appear as they are written, with citations and the support score following.
Scope a question
Limit a question to a document type, a filter or a project, or act on behalf of a specific end user’s permissions.
Harder questions
Agentic mode
The system plans several searches, reads what it finds and decides what to look up next, within a fixed step budget. You see the whole plan, and every step respects permissions.
Longer context for lists
“List all” questions retrieve more passages automatically, so lists come back complete.
Retrieve-only mode
Get the passages without a written answer when your own application writes the response.
How an answer is built
Retrieve
The best passages are found under the asker’s permissions.
Compute
Totals over tables are calculated exactly.
Write
The model writes an answer using only those passages.
Check
Citations are validated and the answer is scored against its sources.
Decline or deliver
Unsupported answers are declined; supported ones are returned with sources.
Questions people ask
Does troveGEN hallucinate?
Language models can, which is why troveGEN limits the model to retrieved passages, validates every citation, scores each answer against its sources, computes numbers from tables, and declines questions the documents cannot answer. These checks reduce the risk; they do not make it zero, so the support score is shown with each answer.
Can troveGEN calculate totals from my documents?
Yes. Sums, averages, counts, minimums and maximums over tables, such as bank statements and price lists, are computed from the extracted rows and shown with the rows used.
Which models can write the answers?
Any OpenAI-compatible model, including self-hosted ones such as vLLM or Ollama, and Azure OpenAI. You can choose per project or per message.
Is the chat private?
Yes. Each person’s threads are visible only to them. Workspace administrators can see how much each person uses chat, not what they asked.
Last updated 2026-10-26. Figures are from troveGEN’s own tests on the dates given; ask us for the method.
Ask your documents something hard.
Start free and put your own questions to your own files.