Your company's own private AI model. ent learns from Confluence, Jira, Slack, Google Drive, Microsoft 365 and your internal systems, runs on-premise on your hardware, and answers with sources.
The answers your team needs are spread across wikis, tickets, chat threads and shared drives, and the people who know where to look are busy. General-purpose assistants don't know your company, and connecting them means sending your data to someone else's servers.
ent builds a model that knows your company and never leaves it. It is trained on your own documents, deployed inside your network, and owned by you: the weights, the index and the data.
Answers in seconds,
with the documents they came from.
Sign in to each system and choose the spaces, projects, channels and folders to include.
Documents are cleaned, secrets and personal data redacted, then turned into training questions by a local model.
An open-weight Qwen model is fine-tuned on your company's knowledge with LoRA, on your own hardware.
Employees ask in plain language. The model retrieves the relevant documents and answers with citations.
Every answer is grounded in your documents and cites them, so people can check the source instead of trusting a guess. The model is trained to point to the document it used.
Open an ITOPS ticket and ask your manager to approve it. Once approved, run acme login staging to get kubeconfig credentials, valid for 12 hours. [1]
A fine-tuned model alone memorizes facts unreliably, and facts go stale. ent combines both: the model learns your company, and a private search index keeps it exact and current.
Fine-tuning teaches your products, teams, terminology and how answers should sound. It travels with the model, no lookup needed.
Retrieval supplies exact figures, dates and the latest version of a document, and shows where each answer came from.
| Source | Model | Index |
|---|---|---|
| Wikis, handbooks, engineering docsCompany-wide knowledge | ||
| Contracts, CRM, HR casesSensitive or fast-changing | ||
| Support transcripts, style guidesHow your company writes |
You decide per source. Retrieval-only data never enters training, so it can be removed at any time and never leaks through the model's weights.
Connections use each vendor's official sign-in, including the Model Context Protocol where available, and ent only ever reads. Content can be limited by space, project, channel, folder and date. Custom systems can be added as connectors.
On a sample company, the trained model with retrieval gave the answers closest to the reference, and grounded answers recalled far more key facts than the base model on its own.
19 held-out questions about a fictional company (wiki, Slack, Confluence and Jira), Qwen3-1.7B on a 16 GB MacBook Pro after 300 training steps. Without retrieval, small models still invent figures; larger models and longer training close more of the gap. Your results depend on your data, which is why every pilot starts with an evaluation on it.
| Runs where your data is | Training, search and answering all run on your machines: a Mac, a GPU server or your private cloud. |
| No hosted AI | Training questions are written by a local model. Sending data to an outside AI service is blocked unless you allow a host you control. |
| Redaction first | API keys, passwords, tokens, card numbers and phone numbers are removed before anything is stored or trained on. |
| Read, never write | Connectors only read from your systems. Credentials stay on your machine, outside configuration files. |
| Control per source | Mark sensitive sources as retrieval-only so they never reach the model's weights. |
| You own it | The trained model, its index and every intermediate file stay yours to keep, audit or delete. |
ent trains and serves on hardware you already have. The model size follows the machine; the workflow stays the same.
| Hardware | Model | Good for |
|---|---|---|
| Mac with Apple Silicon, 16 GB | Qwen3-1.7B | Pilots and single teams |
| Mac with Apple Silicon, 32 GB+ | Qwen3-4B to 8B | Departments |
| One GPU server, 24 to 48 GB | Qwen3-8B to 14B | Company-wide assistant |
| Multi-GPU or private cloud | Qwen3-32B | Large organizations |
The finished model serves an OpenAI-compatible API, so existing internal tools can use it. Retraining on fresh data is one command, so it can run on a schedule.
Yes. ent runs entirely on your own infrastructure. It reads your systems, writes its training questions with a local model, and fine-tunes an open-weight model on your hardware. Sending data to a hosted AI service is blocked unless you explicitly allow a server you control.
A retrieval-only chatbot looks up documents and asks a general-purpose model to summarize them. ent also fine-tunes the model on your company's knowledge, so it understands your products, teams and terminology, and it keeps a private search index for exact, current answers that cite their sources.
ent fine-tunes open-weight Qwen3 models, from 1.7B to 32B parameters depending on your hardware. The trained model, the search index and all intermediate data belong to you.
A pilot runs on a single Apple Silicon Mac with 16 GB of memory. Company-wide deployments use a Mac with more memory, a GPU server, or your private cloud.
Secrets, card numbers and phone numbers are redacted before anything is stored. Anything trained into a model can be surfaced to everyone who uses it, so sensitive sources such as contracts, CRM or HR data can be marked retrieval-only: they are searchable and cited, but never enter the model's weights.
Confluence, Jira, Slack, Google Drive and Docs, SharePoint, OneDrive and Microsoft Teams, plus SQL databases, internal APIs, CSV and JSON exports, and any server that speaks the Model Context Protocol (MCP). Custom systems can be added as connectors.
ent is in early access. A pilot connects a few of your sources, trains a first model on your hardware, and measures it on questions your team actually asks.