Knowledge Base

The knowledge base lets you upload documents and make them retrievable inside workflows. It is the basis for retrieval-augmented generation (RAG).

Collections

Documents are organised into collections. Each collection is a self-contained scope for retrieval. Typical patterns:

one collection per product or domain,
one collection per audience (internal vs. customer-facing),
one collection per language.

Workflows reference collections by name when they perform retrieval.

Supported file types

The Knowledge Base accepts the following formats today:

Text and prose: .pdf, .docx, .pptx, .xlsx, .csv, .html, .md, .adoc, .tex
Images: .png, .jpg, .jpeg, .tiff, .tif, .bmp, .webp

Maximum file size: 50 MB per upload.

Upload flow

Open Knowledge Base, choose or create a collection.
Click Upload, pick a file, fill in metadata.
Pick a chunking strategy (semantic or fixed-size). Semantic is the default and works well for prose.
Submit. The file enters the pipeline.

Behind the scenes:

The original file is stored securely inside your environment.
A knowledge source record is created with status Pending.
A processing job is queued.
A background worker picks up the job, extracts the document content, splits it into searchable chunks, and registers it in the knowledge index.
The status moves through Processing to Completed (or Failed if any step errors).

The UI polls and reflects status changes; you can also refresh manually if you prefer.

Chunking strategies

Semantic - boundaries are inferred from the content structure (headings, paragraphs, list breaks). Best for natural-language documents.
Fixed-size - fixed character windows with configurable overlap. Useful when documents are uniform or when you want predictable retrieval costs.

Both strategies preserve metadata (source title, document URL, page) so retrieval results can be cited back.

Workflow nodes that retrieve from a collection accept a query, run a similarity search, and return the top matching chunks with their metadata. The chunks are then bound into the prompt context for the downstream model node.

Recommended pattern: pass the query through, retrieve the top three to five chunks, inject them with a citation marker, and ask the model to ground its answer in the retrieved context.

Updating and re-indexing

Replace a file - upload a new version. The old source remains until you delete it.
Delete a source - removes the original file, the knowledge source record, and the corresponding entries from the knowledge index.
Delete a collection - removes everything: the collection, all sources, and all original files.

Operational notes

Throughput - processing time scales with document size and the chunking strategy. Large PDFs (50 MB) typically take a couple of minutes; smaller files complete in seconds.
Failures - failed sources surface the underlying error message. Common causes: corrupted PDF, unsupported encoding in CSV, empty document.
Data residency - documents and their searchable chunks live exclusively inside your environment. Embedding generation runs on the local AI engine unless your administrator has explicitly enabled an external embedding model.

Collections

Documents are organised into collections. Each collection is a self-contained scope for retrieval. Typical patterns:

one collection per product or domain,

one collection per audience (internal vs. customer-facing),

one collection per language.

Workflows reference collections by name when they perform retrieval.

Upload flow

Open Knowledge Base, choose or create a collection.

Click Upload, pick a file, fill in metadata.

Pick a chunking strategy (semantic or fixed-size). Semantic is the default and works well for prose.

Submit. The file enters the pipeline.

Behind the scenes:

The original file is stored securely inside your environment.

A knowledge source record is created with status Pending.

A processing job is queued.

A background worker picks up the job, extracts the document content, splits it into searchable chunks, and registers it in the knowledge index.

The status moves through Processing to Completed (or Failed if any step errors).

The UI polls and reflects status changes; you can also refresh manually if you prefer.

Chunking strategies

Semantic - boundaries are inferred from the content structure (headings, paragraphs, list breaks). Best for natural-language documents.

Fixed-size - fixed character windows with configurable overlap. Useful when documents are uniform or when you want predictable retrieval costs.

Both strategies preserve metadata (source title, document URL, page) so retrieval results can be cited back.

Retrieval

Recommended pattern: pass the query through, retrieve the top three to five chunks, inject them with a citation marker, and ask the model to ground its answer in the retrieved context.

Updating and re-indexing

Replace a file - upload a new version. The old source remains until you delete it.

Delete a source - removes the original file, the knowledge source record, and the corresponding entries from the knowledge index.

Delete a collection - removes everything: the collection, all sources, and all original files.

Operational notes

Throughput - processing time scales with document size and the chunking strategy. Large PDFs (50 MB) typically take a couple of minutes; smaller files complete in seconds.

Failures - failed sources surface the underlying error message. Common causes: corrupted PDF, unsupported encoding in CSV, empty document.

Data residency - documents and their searchable chunks live exclusively inside your environment. Embedding generation runs on the local AI engine unless your administrator has explicitly enabled an external embedding model.