Knowledge Base
A knowledge base is a store of your own content, such as product docs, FAQs and policies, that agents search while they talk to customers. When a customer asks something your content covers, the agent looks up the closest passages and answers from them instead of guessing.
This approach is called retrieval-augmented generation (RAG).
How it works
- You create a knowledge base in a project and add content: files, pasted text or a website URL.
- VoisX splits each document into chunks and turns every chunk into an embedding, a numeric fingerprint of its meaning.
- You attach the knowledge base to an agent.
- During a conversation, the agent searches the knowledge base when a question needs it, and the closest chunks are given to the agent before it answers.
The search is semantic. It matches meaning, not exact words, so "how do I connect my calendar" finds a page titled "Calendar setup".
Create a knowledge base
- In your project, open Knowledge Base (under Connect in the sidebar).
- Click New knowledge base.
- Give it a name, and optionally a description of what it holds.
A project can have several knowledge bases, for example one per product line or language.
Add content
Open the knowledge base. On the Documents tab there are three ways to add content.
Upload files
Click Upload files, then drop files on the upload area or browse for them. You can upload several files at once.
| Format | Notes |
|---|---|
PDF (.pdf) | Text is extracted from every page. Scanned PDFs that contain only images have no text to extract. |
Plain text (.txt) | Used as-is. |
Markdown (.md) | Good for exports from docs tools and wikis. |
Other file types are rejected with "Unsupported file type".
Add text
Click Add text, give the document a title and paste the content. This is useful for FAQs, policy wording or anything you don't have as a file.
Ingest a URL
Click Ingest URL and enter a web address that starts with http:// or https://. VoisX then:
- Looks for a
sitemap.xmlon that site to find its pages. If there is none, it follows links from the page you entered. - Fetches up to 80 pages from the same site.
- Keeps the readable text and drops navigation, footers and scripts.
- Chunks and embeds every page.
For a documentation site, enter its root URL (for example https://docs.example.com). The sitemap gives VoisX every page without you listing them.
Pages must be publicly reachable without signing in. Content that a page only renders with JavaScript after it loads may not be captured. If a page comes out empty, add that content as a file or as text instead.
Document status
| Status | Meaning |
|---|---|
| Processing | The content is being chunked and embedded. This usually takes seconds to a minute; a large website takes longer. |
| Active | Indexed and searchable by agents. |
| Failed | Ingestion failed. Check that the URL is publicly reachable or that the file contains readable text, then try again or delete it. |
Keep content up to date
- URL sources: click Refresh on the document. VoisX fetches the site again and replaces the old chunks.
- Files and text: delete the document and add the new version.
Deleting a document removes it and its embeddings from the knowledge base. Deleting a knowledge base removes all its documents, and agents that used it lose access to it.
Attach it to an agent
An agent uses one knowledge base at a time. To attach one:
- Open the knowledge base, go to the Agents tab and assign it to an agent, or
- Open the agent and choose the knowledge base there.
Once attached, the agent gets a built-in knowledge search tool and decides on each turn whether to use it. You don't need to configure the tool.
Answer only from your content
By default, an agent can also answer from the AI model's general knowledge. To stop that, open the agent and turn on Restrict to Knowledge Base & Tools only. The agent then answers only from its knowledge base and tools, and says it doesn't have the information when neither covers the question.
Guide the agent in a skill
You can tell the agent when to search in a skill or in its system prompt, for example:
Search your knowledge base before answering product questions.
If you find the answer, share it and ask whether it solves the problem.
Only offer to open a support ticket if the knowledge base doesn't cover it.
Test retrieval
On the knowledge base's page, use Test retrieval. Type a question your customers might ask and click Search. You see the chunks the agent would get, each with a match score. If the right passage doesn't come back, adjust the content or the settings below.
Then try the agent itself in the Playground.
Retrieval settings
The Settings tab controls how documents are chunked and how many results the agent gets. The settings apply to every document in that knowledge base.
| Setting | Range | Default | What it does |
|---|---|---|---|
| Chunk size | 100–4000 characters | 1000 | Length of each chunk. Smaller chunks are more precise; larger chunks carry more context. |
| Chunk overlap | 0–500 characters | 200 | Text shared between neighbouring chunks, so a sentence split across two chunks isn't lost. |
| Top-K results | 1–20 | 5 | How many chunks the agent gets per search. |
| Similarity threshold | 0–1 | 0 | Drops results below this match score. 0 keeps all results. |
| Embedding model | OpenAI text-embedding-3-small, text-embedding-3-large or text-embedding-ada-002 | text-embedding-3-small | The model that creates embeddings. |
If you change the embedding model or the chunk settings, content you already added keeps its old embeddings and chunks. Re-add your documents (or refresh URL sources) so everything matches.
Usage and billing
Knowledge base searches and storage are metered usage. Some plans may not include knowledge base storage. See Usage and billing and voisx.ai/pricing.
Best practices
- Keep documents focused. A specific FAQ file retrieves more precisely than one large document covering everything.
- Use clear structure. Headings, short paragraphs and lists chunk more cleanly than dense prose.
- Prefer URL sources for live docs, and refresh them when your docs change.
- Remove outdated content. Old answers can mislead the agent.
- Test with real questions in Test retrieval and the Playground after every significant change.