Knowledge Base

Overview

The Knowledge Base lets you upload documents and content so your agents can search and reference them when answering questions. Instead of relying solely on general training data, an agent with a Knowledge Base can find specific information from your own files and give more accurate, grounded answers.

There are two ways an agent can retrieve knowledge:

RAG (Retrieval-Augmented Generation): finds the most semantically similar passages to the user’s question and injects them into the agent’s context.
GraphRAG: builds a knowledge graph from your documents — extracting entities and their relationships — and traverses it to answer complex, multi-hop questions that plain similarity search may miss.

Both can be active at the same time and complement each other.

Documents

Uploading Documents

You can upload one or more files at a time. Supported formats include PDF, Word documents, plain text, HTML, and many others — text is extracted automatically from the file contents.

During upload you can:

Assign tags to the documents for later filtering
Choose whether to replace an existing document with the same name

Once uploaded, a document is queued for processing. The document record is created immediately; embedding and graph creation happen asynchronously in the background.

Uploading Raw Content

Instead of a file, you can push plain text directly — useful when the content comes from a database, external API, or any programmatic source. You provide a name and the text body; the rest of the pipeline is identical to a file upload.

Searching & Filtering Documents

The document list supports:

Full-text search by document name
Filtering by tag
Pagination for large libraries

Each row shows the document name, assigned tags, and the current processing status for embeddings and graph creation.

Updating Documents

You can rename a document or change its tags at any time. These changes do not trigger reprocessing.

Deleting Documents

Deleting a document removes the file, its vector embeddings, and its knowledge graph nodes and relationships. This action cannot be undone. Bulk deletion is supported.

Reprocessing Documents

If you change the embedding model or graph settings after a document was already processed, you can reprocess it selectively:

Reprocess embeddings — deletes existing vector chunks and re-runs the embedding pipeline with the current settings.
Reprocess graph — deletes existing graph nodes/relationships and re-runs entity extraction.

Both options can be combined or run independently.

Processing Pipeline

Text Extraction

When a file is uploaded, text is extracted from its contents automatically. The raw text is then passed to the rest of the pipeline.

Chunking

The extracted text is split into overlapping chunks of a configurable size (measured in tokens). Chunks that are too short are discarded. Each chunk keeps metadata about which document it came from and its position in the document.

Embeddings (RAG)

Each chunk is converted into a vector embedding using the configured embedding model. These vectors are stored in a vector database indexed for fast similarity search. When an agent queries the knowledge base, the user’s message is embedded the same way and the most similar chunks are retrieved and injected into the prompt.

Optional enrichment can be applied to each chunk before embedding:

Keywords — a short list of keywords is extracted and prepended to the chunk text, improving retrieval precision.
Summary — an LLM-generated summary of the chunk (and optionally its adjacent chunks) is prepended, helping the model understand context.

Knowledge Graph (GraphRAG)

When graph creation is enabled, each chunk is also analysed by an LLM to extract entities (people, places, concepts, organisations, etc.) and relationships between them. The result is stored as a graph database.

At query time, the user’s question is used to search for relevant entities by name. The graph is then traversed to gather related nodes and relationships, which are injected into the agent’s prompt alongside or instead of vector search results.

GraphRAG is especially useful for:

Questions that require following chains of relationships (“Who works with whom?”, “What depends on what?”)
Fact-checking across multiple documents
Structured domain knowledge (e.g., ontologies, regulatory frameworks)

Processing Status

Each document tracks two independent statuses — one for embeddings and one for the knowledge graph:

Status	Meaning
Not started	The pipeline has not begun yet
Processing	Currently running
Completed	Finished successfully
Failed	An error occurred — a reason is provided
Skipped	Processing was skipped (e.g., document already up to date)

Token usage is tracked per document for billing and observability purposes.

Settings

Embedding Settings

Setting	Description
Embedding model	The model used to create vector embeddings (e.g. `text-embedding-3-small`)
Chunk size	Target number of tokens per chunk (default: 450)
Min chunk size	Minimum character length — shorter chunks are discarded
Max chunks	Hard cap on the total number of chunks per document
Top K	How many chunks to retrieve per query
Distance type	Similarity metric used for retrieval: Cosine, Euclidean, or Manhattan

Changing the embedding model or chunk size requires reprocessing existing documents to take effect.

Enrichment Settings

Setting	Description
Keywords enabled	Extract keywords per chunk and include them in the embedding
Keywords per chunk	Number of keywords to extract (default: 5)
Summary enabled	Prepend an LLM-generated summary to each chunk before embedding
Adjacent summary	Also include summaries from the preceding and following chunks

Graph Settings

Setting	Description
Graph creation enabled	Whether to run the GraphRAG extraction pipeline on new documents
LLM model	The language model used to extract entities and relationships
Fulltext index	Whether to create a fulltext index on entity names for fast search

Using Knowledge in Agents

To give an agent access to your knowledge base:

Make sure at least one document has been successfully processed (status: Completed).
On the agent’s configuration, enable RAG and/or GraphRAG.
Specify which tags the agent should retrieve from, to scope the search to the relevant documents.

The agent will then automatically search the knowledge base for every user message and include relevant passages in its context before generating a response.

Inspecting Processed Knowledge

Embeddings

You can view the chunks and their embeddings stored for a document. This is useful for verifying that the document was chunked and enriched as expected, and for debugging retrieval quality.

Knowledge Graph

You can inspect the graph extracted from a document:

Graph view — full list of nodes and relationships
Entities view — entity list with types and properties
Graph status — processing metadata including token usage and timing

Limits & Constraints

Maximum file size per upload: 200 MB
Maximum chunks per document: 10,000
Processing is asynchronous — large documents may take several minutes
Reprocessing deletes existing embeddings or graph data before re-running
Changing the embedding model requires reprocessing all affected documents
Deleting a document is permanent and cannot be undone