Features
Knowledge in, grounded answers out, real actions when you're allowed — and the transparency to trust all of it.
Knowledge
Multi-source ingestion
Upload PDFs and Word docs or paste URLs. Everything is parsed, split into meaningful passages and indexed automatically.
Live ingestion pipeline
Watch every document move through extract → chunk → embed → index with real progress — not a spinner.
Document library
One searchable place for every source, with status, chunk counts and previews of what the assistant sees.
Upload
PDFs, Word docs, URLs
handbook.pdf · 4 pages
Extract
Text + metadata per page
pages → text · tenant tagged
Chunk
Overlapping passages
~800 tokens · overlap kept
Embed
Text becomes vectors
chunks → embeddings, in parallel
Index
Into your private namespace
Pinecone · your tenant only
Intelligence
Grounded Q&A
Answers are built strictly from retrieved passages of your own knowledge. If it isn't there, you get “I don't know” — not fiction.
Agentic reasoning
The agent plans: search internal docs? check the web? take an action? answer directly? You watch it choose in the live trace.
Memory
Follow-ups resolve naturally within a conversation, and durable facts carry across sessions — visible and deletable.
Citations you can open
Every fact carries a numbered citation. Click it: the source opens at the exact page, with the grounding passage in a side panel.
Agent trace
- Planning
- Knowledge search
- Drafting
Actions
Pluggable connections (MCP)
Your ticketing, HR and analytics systems connect over the open Model Context Protocol. New integration = new server URL, not new code.
Role-based actions
“Assign the login-bug to Priya and show her workload” — done, if you're a manager. Employees never even see those capabilities.
Employee sessionask & draft
tools discovered at session start
- search_knowledge
- web_search
- action tools — never connected, so never visible
Manager sessionask, draft & act
+ Company MCP Server connected for this role
- search_knowledge
- web_search
- assign_ticketvia MCP
- list_ticketsvia MCP
- get_analyticsvia MCP
Same agent, same code — the role decides whether the Company MCP Server is even connected. Absent by construction, not blocked by a prompt.
Trust & operations
Multi-tenancy
Companies are isolated at the storage layer — each workspace's knowledge lives in its own namespace, always.
Guardrails
Grounding rules, role limits, token ceilings and rate limits — constraints that hold by construction.
Evaluation
A golden dataset and automated judges score faithfulness and retrieval quality on every change.
Analytics & observability
Cost, latency, token usage, quality trends and per-request traces — the numbers behind every answer.
0.000
faithfulness
0.000
answer relevance
0.000
retrieval hit-rate
0
golden questions
Latest production evaluation run — browsable in the product, not a marketing statistic.
See it for yourself
The full product surface is live today — walk through every screen and see how your company's knowledge becomes answers.