Knowledge bases¶

A knowledge base is a set of documents split into chunks and embedded, so a Retriever node can search them.

Before you begin¶

• The admin or editor role (resources.manage). Viewers cannot create knowledge bases.
• An embedding model. Without a credential, embeddings use mock:512, an offline model that matches on shared words, not on meaning.

Create one¶

1. Open Settings → Knowledge bases and choose New knowledge base.
2. Give it a name and choose Create. Settings on the card picks the Embedding model and the Embedding key (a credential).
3. Upload document accepts .txt, .md and .pdf up to 10 MB each.
4. Use Test a search on the card to check what a query returns.

Embeddings¶

Embeddings use the knowledge base's credential, then OPENAI_API_KEY, then a fixed offline mock (mock:512). The default OpenAI model is text-embedding-3-small. For Ollama the default local model is nomic-embed-text (ollama pull nomic-embed-text). Anthropic has no embeddings API.

Test embedding on a credential reports the vector size. Each knowledge base stores its dims. A vector of another size fails with KB_EMBEDDING_DIMENSION_MISMATCH; re-index the base to adopt the new size. After changing the model or the key in Settings, use Re-index.

Storage and search¶

Only the owner can use a knowledge base in a run. Error codes are KB_* in Error codes.

Use it in a graph¶

Start → Retriever (query: {{state.message}}, top_k: 4, output: context)
      → LLM (prompt: Answer {{state.message}} using {{state.context | json}})
      → End

The RAG Q&A starter template has this shape.