Knowledge Bases
Knowledge Bases give your agents long-term, searchable reference material. You upload documents (or ingest web pages), Nexora splits them into chunks and embeds each chunk as a vector, and agents retrieve the most relevant chunks at run time using semantic search — a Retrieval-Augmented Generation (RAG) workflow.
Each knowledge base is scoped to your organization and lives at /knowledge-bases.
Finding Knowledge Bases
Open the Knowledge Bases entry in the workspace sidebar (book icon, near Tools and MCP Servers). The page header shows the total number of bases and documents, a search field to filter by name or description, and a New Knowledge Base button.
When you have none yet, the page shows an empty state with a Create your first knowledge base button.
Creating a knowledge base
Open the create dialog
Click New Knowledge Base (or Create your first knowledge base in the empty state).
Name it
Enter a Name (required, e.g. “Product Docs”) and an optional Description describing what the base contains.
(Optional) Configure chunking
Click Show chunking options to expand the advanced panel and pick a chunking strategy, chunk size, and overlap. If you skip this, the base is created with the defaults (fixed strategy, 512-character chunks, 50-character overlap). Chunking can always be changed later from the base’s detail page.
Create
Click Create. The base appears in the list. Open it to start adding documents.
Uploading documents
Open a knowledge base to reach its detail page, then add content two ways:
File upload
Click Upload and choose a file. Nexora ingests it in the background — the file row moves through Pending → Processing → Ready (or Error if extraction fails). Once Ready, the row shows how many chunks were produced.
Supported file types
| Category | Extensions |
|---|---|
| Documents | .pdf, .docx, .txt, .md, .rst, .html |
| Code | .py, .ts, .tsx, .js, .jsx, .go, .rs, .rb |
| Data | .json, .yaml, .yml, .csv |
PDFs and Word documents have their text extracted automatically; everything else is read as UTF-8 text.
Each uploaded file must be 50 MB or smaller. Unsupported file types are rejected before upload.
You can delete an individual document at any time; its chunks are removed with it. Deleting the whole knowledge base removes all of its files and chunks.
Chunking strategies
Documents are split into smaller chunks before embedding. The strategy and size determine how text is divided, which affects retrieval quality. Configure it per base in the Chunking Strategy panel on the detail page (or in the create dialog’s advanced options).
| Strategy | How it splits | Best for |
|---|---|---|
| Fixed (default) | By raw character count, with overlap between consecutive chunks | Uniform text where structure doesn’t matter |
| Sentence | On sentence boundaries (. ! ?), packing whole sentences up to the chunk size | Prose and documentation |
| Paragraph | On blank-line (double-newline) boundaries, packing whole paragraphs up to the chunk size | Structured docs with clear paragraphs |
Size and overlap
- Chunk size — the maximum chunk length in characters. Range 64–2048, default 512.
- Overlap — how many characters carry over from the end of one chunk into the start of the next, so context isn’t lost at boundaries. Default 50. Overlap must be at most half the chunk size; the slider caps it automatically.
For the sentence and paragraph strategies, an oversized single sentence or paragraph falls back to fixed splitting so no chunk exceeds the size limit.
Changing the chunking configuration affects documents ingested after the change. Re-upload existing documents if you want them re-chunked with the new settings.
Semantic search in chat
Once a base has Ready documents, agents can retrieve from it during a conversation using the built-in knowledge_search tool. Given a natural-language query, it embeds the query, compares it against the stored chunk vectors, and returns the most relevant chunks (ranked by cosine similarity) for the agent to ground its answer on.
The tool is always available to agents — no per-agent configuration is required. Just ask a question whose answer lives in your documents, and the agent searches the org’s knowledge bases automatically.
Embeddings are generated through the organization’s first active OpenAI-compatible provider (OpenAI, OpenRouter, Groq, Together, or Gemini API) using the text-embedding-3-small model. If no such provider is configured, Nexora falls back to keyword-overlap scoring — search still works, but with lower relevance than true vector search. Add an OpenAI-compatible account under Settings → Accounts for the best results.
You can also test retrieval directly: the base’s search endpoint returns the top matching chunks for a query, ranked by score, which is handy for verifying that ingestion produced useful chunks.
Related
| Page | What it covers |
|---|---|
| Settings | Connect an OpenAI-compatible provider for embeddings |
| Agents | Configure the agents that search your knowledge bases |
| Chat Interface | Where agents answer using retrieved context |