Knowledge Base

The Knowledge Base is the repository of information the assistant consults when generating responses. It implements RAG (Retrieval-Augmented Generation): documents are converted into semantic vectors and, for each incoming message, the most relevant excerpts are retrieved and injected into the LLM context.

How it works

Document uploaded
      ↓
Text extraction and vectorization
      ↓
Storage in vector database
      ↓
User sends a message
      ↓
Semantic similarity search
      ↓
Relevant excerpts injected into prompt
      ↓
LLM responds based on the excerpts

Adding content

Go to Assistant → Knowledge and choose one of three ways to add content:

File upload

Drag a file to the upload area or click to open the file picker. Supported formats:

Upload progress is shown in real time.

Add website

Extracts and indexes content from a URL via web scraping.

Use Recursive crawl carefully — it can index hundreds of pages and significantly increase processing time.

Add text

Paste or write content directly to be indexed, without needing a file.

Processing status

After submission, content goes through the following stages:

Status is checked automatically every 5 seconds while processing is in progress.

Managing files

View content

Click the eye icon to see the extracted text rendered as markdown. Long documents use infinite scroll.

Edit

Allows changing the title and reference link of an already-indexed file. The vectorized content itself cannot be edited — delete and recreate the file to update its content.

Delete

Permanently removes the file and all its vectors. This action cannot be undone.

Using knowledge in your prompt

When CustomPrompt is enabled, use the ${{conhecimento}} variable to control exactly where retrieved excerpts appear in the prompt:

You are a technical support specialist for Company X.
Answer ONLY based on the information below:

=== DOCUMENTATION ===
${{conhecimento}}
=== END ===

If the answer is not in the documentation above, say:
"I couldn't find that information. Can I connect you with our technical team?"

Without CustomPrompt, excerpts are inserted automatically after the main prompt. See Generative Agents for more details on CustomPrompt and ${{conhecimento}}.