NOTEGEN.Knowledge base
Build vector indexes for workspace Markdown and retrieve relevant excerpts in conversations.
The knowledge base uses RAG: split Markdown into chunks, create vectors with an embedding model, retrieve similar chunks for a query, optionally rerank them, and add them to Chat or Agent context.
Models
- Embedding (required): add an
embeddingmodel under Model Configuration and select it here. - Rerank (optional): use a
rerankmodel to improve candidate ordering.
Online providers receive text being indexed or candidates being reranked. After changing the embedding model, clear and rebuild vectors.
Build the index
- Calculate vectors manually from file or folder menus.
- With automatic calculation enabled, saving Markdown updates its index.
- Exclude a file or disable automatic calculation for it.
- Folder statistics show index coverage. The Skills directory is not indexed as ordinary content.
Parameters
| Parameter | Range | Effect |
|---|---|---|
| Chunk size | 100–5000 | Larger chunks contain more context but cost more and may reduce precision |
| Overlap | 0–500 | Preserves boundaries but repeats more content |
| Result count | 1–10 | More results add context and possible noise |
| Similarity threshold | 0–1 | Higher is stricter and may return nothing |
Use defaults when unsure. Rebuild affected vectors after changing chunk parameters.
Use in Chat or Agent
Enable Knowledge Base in the conversation toolbar. Source details help verify matched files and excerpts. If nothing is retrieved, confirm indexing, use query terms related to the notes, lower the threshold, or increase result count.
Clear Knowledge Base removes local vectors only; it does not delete Markdown files.