Knowledge bases (Studio)

This page is the UI walkthrough. For the technical underpinnings (embeddings, chunking, search modes), see Features → Knowledge bases.

URL: /app/notebooks (Knowledge Bases live alongside Notebooks under the same shell).

Creating a KB

1. Studio → Knowledge bases → New.
2. Pick a name + visibility:
   • Private — just for you.
   • Group — multi-select groups.
   • Org — every agent in the org can use.
3. Pick the storage mode:
   • Local (Community+) — fast for ≤50 docs.
   • Vector (Pro+) — for >50 docs and multi-user.
4. Pick the embedding model — default bge-m3. You can switch later (re-ingest cost applies).
5. Click Create.

Adding documents

Three ways:

Source	UX
Upload	Drag-drop PDF / Markdown / TXT / DOCX / HTML / EML / CSV.
URL	Paste a URL — Bee Flow fetches and ingests.
Nextcloud folder	Pick a folder; ingestion is recursive. Tracks the folder for future re-ingest.

Ingestion runs in the background. Each document shows a progress bar with chunk count + embedding status.

Browsing chunks

Click a document to see its chunks. Each chunk shows:

• The text
• A "found in N searches" counter
• A button to re-embed just that chunk (after edits)

Attaching a KB to an agent

In the Agent Designer → Knowledge tab:

• Auto-search — KB is searched on every turn.
• Tool-mediated — KB is exposed as a kb_search tool the agent calls when needed.

Auto-search is faster but always retrieves; tool-mediated is more efficient (no retrieval on small-talk turns) but adds one round-trip for the model to decide to search.

Pinning a KB at chat time

@kb:product-docs what's our refund policy?

The @kb: prefix scopes retrieval to a single KB for that turn. Without it, the agent's auto-attached KBs are searched.

KB Marketplace

Pre-built KBs are installable from the Marketplace tab:

KB	Content
Bee Flow Docs	This entire docs site.
GDPR text	Articles + recitals of the EU GDPR.
Nextcloud Admin Manual	The full NC admin docs.
AI ethics frameworks	OECD, EU AI Act, NIST RMF.

Installs are read-only at first — you can fork them to edit.

Settings

Per-KB settings:

Setting	Notes
Embedding model	Switch model, triggers re-embed.
Chunk size	Override KB_PER_CHUNK_TOKEN_CAP.
Search mode	Vector / BM25 / Hybrid / Reranked.
Reranker	Azure / Cohere / local cross-encoder.
Retention	"Forever" / "auto-delete after N days".

Sharing

A KB shared at org level is visible to every agent in the org. Per-agent attachment is still required — sharing makes it eligible to attach, not auto-attached.

Deleting

Soft-delete (30-day recovery window) or hard-delete (immediate). Deleting a KB removes its chunks but does not delete the original documents stored in NC / Drive — Bee Flow only ever held copies.

Troubleshooting

Symptom	Likely cause	Fix
Ingestion stuck at 0%	Document parser couldn't read the file	Check the format; fall back to text export.
Searches return nothing relevant	KB hasn't ingested yet, or query mode mismatch	Wait for ingestion; switch to Hybrid.
Agent doesn't cite KB	Auto-search off, or kb_search tool not in allow-list	Toggle in Agent Designer.