Demo

Guide · Beginners

Build your first assistant in three steps

Connect a language model, build a corpus, configure the assistant and test it — the basic workflow in EasyRag.

Beginner 15 minUpdated

This guide follows the setup path described in the user documentation: a model credential, a corpus, an assistant. It assumes you have an account with the Administrator or Owner role on an EasyRag instance.

1. Add a language model credential

In the administration area, open Credentials and add a credential:

  • choose the provider (OpenAI, Mistral, Anthropic, Google Gemini, or an Ollama, vLLM or OpenAI-compatible server);
  • enter the API key or, for an internal server, its address;
  • give it a meaningful name (for example “Mistral — production”).

The key is encrypted as soon as it is saved. You can share it with named users without ever revealing it to them.

Use the model test button to check the connection before going any further.

2. Create a document corpus

Open Corpora, create a corpus (for example “HR procedures”) and import your documents.

Supported formats: PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), CSV, plain text, Markdown, HTML and JSON. Images, scanned PDFs, audio and video files require additional keys (OCR, transcription) configured by the instance administrator.

Each document goes through text extraction, chunking, contextual enrichment and vectorization. If content validation is enabled, it may also be quarantined pending a decision.

Use the corpus test search to check that your documents are being found.

3. Create and configure the assistant

Open Assistants and launch the creation wizard:

  1. describe the goal in one sentence (“Answer employee questions about leave and absences”);
  2. the AI suggests a name, a description, a system prompt, a welcome message and starter questions;
  3. choose the credential and model set up in step 1;
  4. select the corpus from step 2.

Review the system prompt: it should specify the role, the scope, the tone and the expected behavior when the documents do not contain the answer.

4. Test

Open the assistant’s test window and ask:

  • a question whose answer is clearly stated in a document — check the citation;
  • an out-of-scope question — the assistant should say so;
  • a question worded differently from the document — the search should still find it.

Every configuration change creates a new version: you can compare versions and roll back at any time.

What’s next?