Ertas for Internal Knowledge Bases

The Challenge

Every organization accumulates institutional knowledge across wikis, Confluence pages, SharePoint sites, Slack threads, email archives, and the minds of long-tenured employees. Finding the right information when you need it is one of the most persistent productivity challenges in modern organizations. Employees spend an estimated 20% of their work week searching for information, asking colleagues questions, and re-creating knowledge that already exists somewhere in the organization.

Generic AI chat tools and enterprise search solutions partially address this problem but fall short in critical ways. They can retrieve documents that contain relevant keywords but cannot answer questions that require synthesizing information across multiple sources or understanding organization-specific context. Questions like 'What is our process for handling a client data breach?' or 'Which team owns the billing integration with Stripe?' require knowledge of the organization's specific processes, team structures, and technology stack, context that generic models do not have and keyword search cannot provide.

The Solution

Ertas lets organizations fine-tune knowledge base models on their own internal documentation, Q&A pairs from help desks and support channels, and process documentation. The resulting model learns the organization's terminology, team structures, tools and procedures, so it answers questions in the company's own context and points employees to the right resources. Combined with RAG over current documents, it answers from both what it was trained on and up-to-date documentation.

In Ertas Studio, knowledge management teams train on curated Q&A pairs drawn from internal support channels, IT help desk tickets, HR FAQ responses and expert interviews. The model learns the context around each answer: which team to contact, which tool to use, which process to follow and which exceptions apply. Exported from Hub and served on internal infrastructure through AnythingLLM, Dify or a custom chat interface, it becomes a conversational assistant employees can query in natural language, with questions and answers staying inside the company network. Training data is uploaded to Ertas's cloud GPUs, so keep restricted material out of the training set. Document-level permissions belong in the retrieval layer. A trained model can repeat anything in its training data to anyone who asks, so train it only on knowledge every user of the assistant is allowed to see.

Key Features

Example Workflow

A 2,000-person technology company has institutional knowledge spread across Confluence (15,000 pages), Slack (3 years of history) and a Zendesk help desk (50,000 resolved tickets). The IT and People Ops teams extract 25,000 Q&A pairs from help desk tickets, Slack support channels and curated FAQ documents, remove personal details and anything restricted to particular teams, and upload them to Ertas Studio. Subject experts review the HR and security answers in Data Craft before approving them. The team fine-tunes a model that knows the company's tools (internal system names, acronyms, team names), processes (onboarding, procurement, incident response) and policies (PTO, expenses, data classification), and compares runs against a held-out set of real employee questions. The chosen model is exported and served through AnythingLLM on internal infrastructure, with Confluence pages loaded as RAG context under existing permissions. Employees reach it through a Slack bot and a web interface. The team tracks help desk ticket volume and answer accuracy against its own evals, and retrains each quarter from newly resolved tickets.