How to create a Virtual Assistant with your own knowledge base
In this guide, you will learn how to create your own Agent – a virtual assistant that answers questions based solely on the documents you upload.
Why create an Agent with your own knowledge base?
Standard AI models use general knowledge from the internet. To work with your documents (e.g., pasted into the chat), they must fit them into their "short-term memory." However, this memory is limited – with long texts, the system gets lost and omits facts. Our system solves this problem. The assistant does not "memorize" the files, but actively searches them.
Main advantages
- Direct control over sources: You decide which documents the artificial intelligence uses.
- Fewer hallucinations (making things up): Answers are more accurate because the assistant is based on hard sources.
- Agentic behavior (autonomy): The model has the freedom to act. It analyzes your files, verifies facts, and searches until it extracts the most important information.
Step 1: Agent Configuration
To get started, go to the Agent Creator tab (robot icon in the side panel).
- From the drop-down list, select the Create a new agent (1) option. Then, by clicking the plus icon (2), you can optionally assign a custom avatar to the agent.
-
Fill in the agent's details:
- Name (3): This is the only required field, enter a specific name here (e.g., "FERS Documentation Assistant").
- Other options (optional): You can fill in the Description (4) and assign the agent to a specific Category (5), which will make it easier to organize in the future.
-
Define the agent's behavior in the Instructions (6) section: Paste the prompt here that determines the agent's role and how it responds. An example instruction can be found in the next subsection. You can use Variables (7) if you want the agent to have knowledge of, for example, the current date.
- Select the appropriate tools:
- Model (8): Choose the model version best suited to your tasks.
- Capabilities (9, 10): Here you can expand the agent's knowledge. Use File Context to upload documents whose content will be permanently attached in full to the prompt, or enable File Search if the agent is to search large sets of documents in real time. We will learn when to use which solution later in this guide.
- Optionally, fill in the Technical Support Contact (11) section so users know who to turn to in case of problems.
- You can expand the Advanced (12) section to adjust additional settings and parameters of the agent (details of this configuration will be discussed in the next guide).
- Finally, confirm everything by clicking the green Create button. After creating the agent, the option to upload documents to it will appear.
- Example form fill:
| Field | What to enter |
|---|---|
| Name | Specifically, e.g., "FERS Documentation Assistant" |
| Description (optional) | Optional, e.g., "Answers questions about recruitment" |
| Category | Leave the default value |
| Instructions | Copy and paste the text block below (you can modify it): You are a helpful assistant answering questions based solely on the provided documents. Information search: Search documents using the file_search function. If the first search did not yield sufficient results, try again with a differently phrased query. Do not perform more than 3 searches per user question. If you cannot find an answer despite your attempts, say so directly. Response style: Answer briefly, specifically, and to the point. Use bullet points and bold text to make the response readable. Do not quote entire document fragments — extract what is essential from them. If the question is too general or you lack context, ask the user before you start searching. Language: Answer in the same language the question was asked. |
| Model | Leave the default selected value |
| Support Contact | Leave blank |
- In the Capabilities section, be sure to check ☑ Enable file search.
- Click the green Create button. Only after creating the agent will the option to add documents to it be unlocked.
View of the filled form.
Step 2: Uploading your knowledge
Once the Agent is created:
- Click the Upload to file search button and select documents
- You can upload a maximum of 50 files at once (you can add more later).
- When finished, click Save. Your Agent is ready!
- Important: Indexing files in the background may take a few minutes. Even when the progress bar disappears, the system is still processing them.
Supported formats:
- Documents: .pdf, .docx, .md, .txt, .epub, .rst
- Data: .xlsx, .xls, .csv, .json, .yaml, .xml
- Code/Logs: .html, .py, .rs, .sql, .sh, .log
.doc and .ppt formats are not supported yet.
View of the panel for adding documents using button (1). After uploading documents to file search, click button (2) to save the changes.
How to talk to your Agent?
- Click the model selection icon at the top of the screen.
- Go to the My Agents tab.
- Select your Agent and start asking questions.
You can now ask questions. The Agent will search your files to provide answers based on the documents.
Model selection menu with the "My Agents" tab and a list of agents
How the Agent thinks in practice?
When you ask it a complex question, you will see messages like "Thoughts" or "file_search launched". This means that the Agent is not guessing, but is taking action:
- It searches the uploaded files for keywords.
- It reads the found fragments.
- If details are missing – it searches again until it collects a complete set of facts.
Multiple searches of the knowledge base
Best practices
The Agent relies on keywords and context. The more precisely you ask, the better it will hit the mark.
- Be very specific: "Can I finance a full 3-year study cycle as part of eligible costs if the project ends on 09/30/2029?"
- Guide it if it gets lost: If the answer is incomplete, try: "Search more precisely, I mean the durability requirements in the regulations, not in the FAQ document."
Limitations of file search and when to choose a different approach
File Search (File Search) works like an intelligent search engine, extracting the best-matching fragments from documents. It does not read the files in their entirety.
What it is good for:
- Questions about specific information: "What is the deadline for submitting applications?"
- Finding definitions and requirements: "What are the eligibility criteria?"
- Large document databases (dozens, hundreds of files).
What it will not handle: - Extracting data from all files at once — e.g., "Collect names and emails from 30 CVs into a table." The Agent will only search a few fragments, it will not process each file separately
- Summarizing the entire database — e.g., "Summarize all uploaded reports"
- Comparing documents — e.g., "How do these three contracts differ?"
File Context: an alternative when you need full content
If you want the agent to see the entire document (not just fragments), use File Context instead of File Search.
| File Search | File Context | |
|---|---|---|
| How it works | Searches and returns only the most important fragments | Loads the full content of the files |
| How many files | Hundreds/thousands | Up to approx. 30–50 short files / ~100-200 A4 pages |
| Best for | Finding specific information (e.g., "Where is invoice X?"). | Cross-sectional analysis and summaries (e.g., "Summarize trends from 20 reports"). |
| Example question | "What is the contractual penalty in paragraph 5?" | "Compare these 10 offers and select the best one, taking all points into account". |
How to add files to Context:
In the Agent editor, in the "File Context" section (above the File Search section), click "Upload file context" and select documents.
You can combine both approaches in one Agent — e.g., a large database of regulations in File Search and a few key documents in File Context.
What about hundreds of files to process? If you have, for example, 500 CVs to analyze, neither File Search nor File Context will handle it — this requires an individual approach. Contact us, we will help you choose the right solution.
Security and trust
- Privacy: All your documents, queries, and search results stay on our servers. We do not send your data to third parties.
- Limited trust principle: AI is only a tool and can make mistakes (so-called hallucinations). Always verify the most important information at the source (in your documents) before making binding decisions.