Choose the scope
Ask across all company knowledge, only internal company knowledge, or one fictional client account.
The Company Knowledge AI
Ask your company what it knows, and get the sources back.
Your team already wrote down most of what the AI needs: emails, meeting notes, PDFs, policies, decisions, contracts, project history. It is scattered across six tools and a few people's heads. This demo turns it into a private assistant that synthesizes an answer in plain language and cites every artifact it used.
Scattered sources
Private knowledge layer
Synthesized, not just searched
Normalized into one company memory, searchable in plain language, and traced back to the exact artifact behind every claim.
Synthesized answer
"SOC 2 readiness is blocked on vendor evidence and audit-log ownership. Sources: security policy, all-hands recap, vendor contract summary."
Demo walkthrough
One question. Every source visible.
Source-backed workflow
01 / Recognition
The operational reason
The answer exists. It is just spread across six tools and four people.
Generic chat tools do not see any of it. Your real context lives in emails, meeting notes, PDFs, internal policies, decision memos, vendor contracts, and project history. This demo shows what happens when that material becomes one private knowledge layer that the assistant can read, synthesize, and cite.
- A new hire asks how escalation works, and the answer is split between the onboarding handbook, an all-hands recap, and a Slack message from last quarter.
- Leadership asks why we picked Postgres and pgvector, and someone has to reconstruct the engineering decision from old meetings and a half-written memo.
- A client call is in 20 minutes, and the open risks are buried across an MSA addendum, a compliance memo, and three meeting notes.
- Procurement, compliance, and engineering all wrote something about the same vendor, and nobody knows which version is current.
02 / Demo moment
Try the live moment
One question. One synthesized answer. Every source visible.
Live demo notes
Fresh notes join the same citation-backed memory.
The public demo also accepts brief visitor-scoped notes. Add one during a session, ask about it immediately, and the assistant treats it like another source it can cite.
Live-note workflow
Ask in plain language
Catch up on a topic, trace a decision, summarize a vendor contract, or prep for a meeting.
Verify every source
Open the cited email, memo, handbook, contract, or transcript and confirm the answer against the original.
Try asking
- Catch me up on SOC 2 readiness.
- Why did we pick Postgres and pgvector?
- What does the onboarding handbook say about escalation?
- Brief me before the Caldrisa Dental call.
- What open risks are flagged across the company right now?
03 / Workflows
Flagship workflows
Built around repeatable synthesis work, not another empty chat box.
Topic catch-up
Get a synthesized brief on what the company knows about a topic: current status, recent updates, open questions, and the artifacts behind each line.
Decision archaeology
Reconstruct what was decided, why, who agreed, and which memo, meeting, or email locked it in. No more "I think we discussed this in March".
Meeting prep
Focus the assistant on one account or one topic, then get a one-page brief on the current state, open risks, and the next questions worth asking.
04 / Privacy
Why private
The useful context is the sensitive context. That changes the architecture.
- The good answers need the sensitive material: contracts, board notes, customer records, salaries, incident write-ups, legal memos. None of that belongs in a generic SaaS chat tool you do not control.
- Deployment is an architecture choice: run on your own servers, in a private cloud, or inside an EU-controlled environment. The reference stack ships with a sovereign vLLM path out of the box.
- Not every step needs a frontier model: extraction, classification, routing, and structured summaries can run on smaller controlled models at a fraction of the cost. Synthesis goes to the larger one only when it earns it.
- Citations are not a feature, they are the audit trail: every answer points back to the email, memo, contract, transcript, or handbook it used, so a human can verify before acting.
- You own the whole stack: database, artifacts, prompts, retrieval, deployment, and model routing all stay under your control. No vendor lock-in on the layer that matters most.
05 / Architecture
How it works
Three stages between scattered files and a verifiable answer.
Ingest the real formats
Emails, PDFs (including scans through OCR), Word documents, Markdown memos, meeting transcripts, and structured exports are parsed into normalized records with metadata.
Index in Postgres
Hybrid retrieval combines full-text search and pgvector embeddings in one database, so every chunk keeps a hard link back to the source artifact.
Synthesize with citations
The assistant builds a context pack from the retrieved chunks, writes a structured answer, and a citation check refuses claims that have no supporting source.
06 / Open source
From demo to real deployment
A working blueprint, not a slide deck. Read the code before you read the contract.
The public repository ships the core pattern end to end: ingest the scattered formats, normalize them into an AI-ready knowledge layer, retrieve with hybrid search, and synthesize answers with citations checked against the source pack.
You can clone it, read every line, and run it locally in one command. That is the point: an open-source reference is a trust anchor for the technical teams, partners, and buyers who want to inspect the moving parts before a sales conversation, not after.
The paid work is everything the public demo deliberately leaves out: real connectors to your systems, per-user permissions, evaluation sets against your domain questions, monitoring, human review steps, and the workflow-specific assistants your team actually needs.
View the open-source reference architecture07 / Fit
Who it is for
Best fit: knowledge-heavy teams where the company's edge lives in what it has already written.
Teams
European mid-market and enterprise teams that need EU-controlled deployment, not a US-hosted SaaS chatbot.
Operations, engineering, compliance, and delivery groups that share context across functions and cannot afford another silo.
Consultancies, agencies, and IT service providers whose intellectual property is the body of documents, decisions, and client history they have accumulated.
Triggers
The same onboarding, policy, vendor, or "why did we do it this way" question gets asked every other week.
A decision needs to be defended to leadership or a regulator, and the trail of reasoning is scattered across old meetings and half-finished memos.
Leadership has decided that pasting sensitive context into ChatGPT is not the long-term plan, and wants a concrete private architecture to evaluate.
08 / FAQ
Questions buyers usually ask
Private AI is a workflow and data-boundary decision, not just a model choice.
Is this just another chatbot wrapper?
No. The chat box is the easy part. The real work is the layer underneath: parsing your real document formats, indexing them in a way that supports hybrid retrieval, building a context pack that fits the model, and refusing answers that cannot be cited. The demo is built around concrete synthesis workflows (topic catch-up, decision archaeology, meeting prep), not a blank prompt.
Why not just use Microsoft Copilot, ChatGPT, or Gemini?
Use them when the data boundary is acceptable, the model choice is acceptable, and the workflow is generic enough. This architecture is for the cases where it is not: sensitive context, EU deployment requirements, an evaluation set you can actually own, and workflows that need to be tuned to how your team actually works.
Does this replace our existing tools?
No. It sits above the systems where the context already lives, like email, documents, notes, ticketing, project systems, and knowledge bases. The goal is to make scattered context usable without forcing a new source of truth on the team.
What would a real deployment add beyond the demo?
Real connectors to your systems, per-user permissions and row-level access control, an evaluation set built from your own domain questions, monitoring and observability, human-in-the-loop review for sensitive outputs, and workflow-specific assistants instead of one generic chat.
Implementation offer
Use the demo to decide where your first private AI workflow should live.
If the pattern fits something inside your onboarding, operations, engineering, compliance, or client-service work, the next step is concrete: map one real workflow, inspect the data boundary, and define a safe pilot scoped to your actual systems and questions. Not a six-month transformation programme, just the first thing worth shipping.