Knowledge Base
A Knowledge Base (KB) is a set of documents the agent can search while talking to a user. Use it when you want the agent to answer from your own content — manuals, FAQs, internal wikis, product documentation — rather than relying solely on the LLM's general knowledge.
Each agent supports exactly one Knowledge Base. Plan your agent architecture accordingly. If different audiences need access to different content, build separate agents.
What you can add
Two content sources, combined in the same Knowledge Base:
- Files — PDF, DOCX, TXT, MD, HTML, and similar document formats.
- Web URLs — the platform crawls and indexes the page (and its linked descendants, depending on scope).
Setting up a Knowledge Base
The Knowledge Base Integration section is always visible in the agent editor — it is not a separate step or a button to click. The form fields are:
- Knowledge Base Name — e.g., "Product Documentation".
- Description — a short note about what the KB contains.
- Add Documents button — opens a file picker to upload documents, or a field to add web URLs.
Fill in the name and description, add at least one document or URL, then click Create Agent (or Update Agent). The platform creates the Knowledge Base and begins ingestion in the background.
Ingestion status
The KB row in the editor shows a status chip after the agent is saved:
| Status | Meaning |
|---|---|
| Setting Up | KB is being created; content is being chunked and embedded. |
| Updating | New content is being processed. |
| Configured | Ready. The agent can answer from this KB. |
| Failed | Something went wrong. Click the row for the error. |
The agent answers from Configured KBs only. KBs still ingesting are skipped until they reach Configured.
You can also monitor KB ingestion progress on the Tool Configuration page — KBs appear alongside remote tools with the same status chips.
How it works
- You add files and URLs to the KB form and save the agent.
- The platform splits each document into chunks, embeds them as vectors, and stores them in a managed knowledge base.
- During a conversation, the agent automatically searches the KB for chunks relevant to what the user asked.
- The most relevant chunks are added to the LLM's context for the reply.
You don't write any retrieval logic — this happens automatically.
Smart defaults
The dashboard uses optimized defaults out of the box:
- Embedding model: Amazon Titan Text Embeddings V2.
- Chunking strategy: Hierarchical (1500-token parents, 300-token children, 60-token overlap).
- Reranking: off.
- Retrieval count: 5 chunks per query.
These defaults work well for the vast majority of use cases. Advanced customization via the JSON editor is available but rarely needed — see Advanced Mode if you need to override them.
Recommended models
When you have a Knowledge Base, the dashboard recommends Anthropic Claude models:
"For better performance with larger Knowledge Bases, we recommend using Anthropic Claude Models."
Claude models handle long retrieved contexts well. Other models work too — Claude is the safe default for KB-heavy agents.
Updating content
- Add new files / URLs and click Update Agent — they are ingested and added to the KB.
- Remove a file / URL and save — that content is no longer searchable.
- Re-uploading a file replaces the old version.
Quotas
Each admin has a maximum number of Knowledge Bases they can own. When you hit the cap:
- The Add Documents area in the form is disabled.
- An error message says something like: "Cannot add more than N knowledge bases per user. You currently have M."
Contact your platform admin to raise the cap.
When NOT to use a Knowledge Base
- Data that changes by the minute. KBs are best for stable content. For live data, use a tool that queries the source directly.
- Very small reference material. A few paragraphs belong in the system prompt.
- Per-user data. KBs are shared across everyone using the agent. For user-specific data, use a tool that scopes to the individual user.
Tips
- Smaller, focused KBs outperform large mixed archives. A 50-page focused manual beats a 5,000-page archive of mixed content.
- Use clean source material. Garbage in → garbage out. Avoid PDFs that are scanned images — use native text formats where possible.
- One agent per audience. If different groups need different content, build different agents.
- Test with real questions. After ingestion, open a chat and ask questions specific to your content. If answers are off, check your source material or system prompt first — retrieval is usually not the problem.