GraphRAG

GraphRAG turns a company's policies, manuals, price lists and catalogs into a knowledge base that an AI assistant answers from. For each question it decides what to open: a document section, a product card by SKU or a selection from the catalog. The answer is built only from what was read, and the list of sources is assembled by the program, so a document the assistant never opened cannot appear under the answer. With the Clients module on, the assistant also books: it offers free time from the real schedule, files a request and reminds the client of the visit.

The product fits where the knowledge already exists but lives in files and in people's heads, and routine questions eat staff time: customer support, sales teams with large catalogs, HR and internal policies, salons, clinics and other businesses that take appointments. Instead of a chatbot with a scenario tree, the company gets an assistant it tunes to its own needs and checks with tests.

The key difference is that answer logic is described as data. Input checks, branches, prompts, the tool set and required phrases live in a harness: JSON with versions, built-in tests and one-command rollback. The harness changes without a service release, and a connected AI agent can edit it over the admin MCP link, while a person switches the new version on. Data stays inside your perimeter: each client or department has its own isolated collection.

A classic bot runs a “router → search → answer” chain: every new type of question needs a new branch. The agent in GraphRAG works differently. It has tools and rules, and it picks the path to the answer itself: for “compare A and B” it opens both cards, for “anything similar?” it takes the product's category and selects by it, for “book me Saturday morning” it checks the schedule and offers slots.

What must always happen is not left to the model. The harness can drop spam without calling the model and switch on a cautious mode for sensitive topics, while required phrases and sources are added by the program.

The harness describes how a collection answers: 12 step types (checks, branches, variables, search in a given section, custom LLM steps, the agent, the answer), prompts, limits and tests. A validator checks it before saving, and the flowchart in the admin panel is drawn from the same data.

Improvement runs in a loop: run the tests, review the answer log, fix a prompt or a step, check the new version on the same questions and switch it on. This can be done in the admin panel or over the admin MCP. A version with failing tests won't go live; rolling back takes one command.