Private Enterprise AI Chat with RAG, Governance, and On-Premise Control

Private enterprise AI chat is a ChatGPT-style assistant where the language model, document retrieval, embeddings, and conversation history all run inside your organisation's perimeter. No prompts, no documents, and no embeddings are sent to an external model provider. VDF AI Chat ships this as a turn-key product with multi-agent orchestration, RAG, code generation, and integrations to Jira, Zoom, and Google/OneDrive — deployable on-premise, in a sovereign cloud region, or fully air-gapped.

Cloud RAG (used by ChatGPT Enterprise, Microsoft Copilot, and most hosted assistants) sends document chunks to a third-party model provider on every query. That creates data residency, IP-leakage, and procurement issues for regulated teams. Private RAG keeps the embedding model, vector database, and generation step inside your environment, so confidential documents never leave the perimeter. VDF AI Chat's RAG layer is private by default: documents are chunked, embedded, and indexed locally, and retrieval is filtered by the requesting user's own permissions before anything reaches the model.

Open-weight models served inside your perimeter — Qwen, DeepSeek, Mistral, and Llama families are the common choices — plus any OpenAI-compatible endpoint. Commercial APIs such as OpenAI, Anthropic, and Azure OpenAI can be enabled per agent where policy allows, and disabled entirely in on-premise and air-gapped deployments. Model choice is a per-agent setting, so a public-data agent and a confidential-data agent can run on different models under different policies within the same workspace.

Retrieval is permission-aware rather than index-wide. Each connected source — Confluence, Jira, GitHub, SharePoint/OneDrive, Google Drive, or an uploaded corpus — carries its source ACLs into the index, and every query is filtered against the requesting user's identity before candidate chunks are scored. A user cannot retrieve a passage from a document they could not open directly in the source system, which is the failure mode that stalls most enterprise chat rollouts.

Every turn writes an append-only record: the requesting identity, the agent and model version used, the retrieved chunk IDs with their source documents, any tool calls made, and the response returned. Because retrieval provenance is captured per turn, you can reconstruct exactly which source passages produced a given answer months later. Logs are exportable to your SIEM and are retained in storage you control, under your own retention policy.

A standard on-premise deployment reaches a working pilot in roughly two weeks: infrastructure and model serving in week one, identity integration and the first indexed knowledge sources in week two. Broader rollout is paced by source onboarding and permission mapping rather than by the platform itself — each additional connected system typically adds days, not weeks. Air-gapped installations add time for the offline model and image transfer.

A departmental pilot of 50–200 users runs on a single GPU node — a 48–80 GB card such as an L40S, A100, or H100 — serving a mid-size open-weight model, alongside modest CPU nodes for the application, vector store, and database. Production deployments in the 1,000+ user range typically run two to four GPU nodes behind a load balancer for redundancy and concurrency headroom. Retrieval and embedding workloads are far cheaper than generation and usually need no dedicated GPU.

Microsoft 365 Copilot is licensed per seat and runs inference in Microsoft's cloud against Microsoft Graph data. VDF AI Chat is licensed as a flat platform fee, runs inference on infrastructure you control, and indexes whatever sources you connect — including systems outside the Microsoft estate. The practical differences that decide most evaluations are inference locality, per-seat versus flat cost at full-company rollout, and whether you can choose and swap the underlying model.

Yes. Most teams adopt it to retire shadow ChatGPT use — chat with any LLM (open or proprietary), drag-and-drop documents into the chat for RAG-backed answers, generate code with sandboxed execution, render HTML artifacts and Mermaid diagrams, and dictate prompts by voice. The difference: the conversation, the documents, and the embeddings all stay inside your governance perimeter, and usage is never fed back into a vendor's model training.

VDF AI Chat can be deployed on-premise, in a sovereign cloud region, or fully air-gapped. All conversation history, file uploads, retrieval indices, and generated artifacts live in storage you control. Audit logs are immutable and exportable. The platform is designed to align with EU AI Act transparency and logging obligations, GDPR residency requirements, and sector-specific rules such as DORA, NIS2, and HIPAA.