Knowledge Base & Assets

Assets are files and data sources your agents can use for retrieval or templating. Upload files, connect data sources, and attach only the relevant assets to the steps that need them — semantic search, not full-document stuffing.

What you can attach

Any source can be attached to one or many agents.

How retrieval works

When you attach a source with vectorization enabled:

1. Fruxon extracts text and splits it into semantic chunks.
2. Chunks are embedded with your chosen embedding model and stored.
3. At run time, the agent's query is embedded; the most relevant chunks are retrieved and injected into the step's context.
4. The agent's response can include the source citations.

You can disable vectorization for a source — useful when you want to reference the document literally (e.g., a template) rather than retrieve excerpts.

Uploading a file

1. Open Assets in the sidebar.
2. Click Upload.
3. Pick a file, name it, optionally tag it.
4. Choose whether to vectorize (default: yes).

The file is processed in the background. Status shows up in the Assets list — Pending → Processing → Completed (or Failed, with an error you can inspect).

Vectorize later

If you uploaded a file with vectorization off (to use it as a literal template), you can vectorize it later without re-uploading. Open the asset, toggle Enable retrieval, and the existing file is re-processed in place — same asset ID, same references, now retrievable. Useful when a template you originally pinned as literal becomes worth searching over.

Connecting a data source

For Confluence, Notion, GitHub, Google Drive, etc.:

1. Make sure the underlying integration is connected at the organization level.
2. In Assets → Add Data Source, pick the integration and the scope (a specific space, repo, folder, page).
3. Fruxon syncs the content and keeps it refreshed on a schedule you control.

Tool sources

A tool source turns a list-style integration tool into a knowledge asset. Where a data source connects a whole space or repo, a tool source pulls records — rows from a database, issues from a tracker, contacts from a CRM — and indexes each one as a document. It rides the credentials of an integration config you've already connected.

In Assets → Add Data Source, pick Tool source and configure:

The field mapping uses dotted paths (e.g. name, fields.summary); a value containing {{…}} is treated as a template. Click Fetch sample record to pull one real record from the tool so you can pick fields by clicking rather than guessing path names. You map:

• Content — the text to embed (the body of the document).
• Document id — each record's stable id, which keys the document and dedupes re-syncs.
• Optionally an image URL and a last-updated path for richer lineage and future incremental sync.

Per-document enrichment

Flat records often don't make great documents on their own. Toggle Enrich documents with an agent to run an agent once per record before it's indexed. The enrichment agent receives the record's content (and any image_url) and must return JSON { embed_text, attributes } — the embed_text becomes what's embedded, and attributes become document metadata. This is where you summarize a raw row into searchable prose, describe an image, or pull structured fields out of messy text.

Indexing from a Collection Pipeline

The reverse direction also exists: a Collection Pipeline with an Index sink writes its per-item results into a knowledge asset as one finalized snapshot. So you can either pull-and-enrich records into an asset (a tool source), or fan an agent out over a data source and publish its outputs to an index (a pipeline) — both land as searchable documents your agents can retrieve.

Attaching to a step

In Studio, select an Agent Step and open its Knowledge Base section. Pick the assets and data sources that step can search. Different steps can attach different assets; a step with no knowledge base has no retrieval cost and cannot search tenant assets by accident.

Referencing assets directly

Sometimes you want the literal contents of an asset, not retrieval. In the prompt editor, open the asset picker and insert the asset — it drops a <file:…> token into the prompt that resolves to the asset's content at run time:

<file:contract_template.docx>

You insert these from the picker rather than hand-typing them, so the reference always points at a real asset. This is common for templates (DOCX, HTML, system prompts) and reference data (CSV, JSON).

Supported file formats

• Documents — PDF, DOCX, TXT, Markdown, RTF
• Data — JSON, CSV, YAML, XLSX
• Web — HTML, URLs (auto-crawled)
• Images — PNG, JPG, WebP, GIF (used as media in document generation)

Best practices

• Chunk-friendly content beats long blobs. Documents with clear headings and short paragraphs retrieve better than walls of text.
• Attach what's relevant, not everything. A focused knowledge base produces better answers than a kitchen-sink one.
• Refresh data sources. For Confluence/GitHub/Drive, set sync schedules so retrieval reflects current state.
• Use citations. Have your agent quote the source — users trust answers they can verify, and you can audit hallucinations faster.

Next steps

• Studio — attach assets to a workflow
• Integrations — connect Confluence, GitHub, Drive, Notion as sources
• Use Cases — RAG and document Q&A patterns