How to access Insights
Click Insights in the left sidebar of the dashboard. The graph loads your current document library and displays it automatically.How the graph is built
Building the graph involves two steps: fetching your vectors from Qdrant, then projecting them into 2D space.Fetch vectors
Prism retrieves up to 100 of your document chunks from Qdrant, including their 768-dimensional embedding vectors and metadata (document name, type, category, chunk index).
UMAP projection
The high-dimensional vectors are reduced to 2D coordinates using UMAP (Uniform Manifold Approximation and Projection). UMAP preserves the local neighborhood structure of the original space — chunks that are semantically close in 768 dimensions end up close in the 2D projection.UMAP parameters:
- nComponents: 2 (output dimensions)
- minDist: 0.1 (how tightly points cluster)
- spread: 1.0 (scale of the embedding)
- nNeighbors: adaptive (up to 15, scaled to the number of points)
Reading the graph
Nodes
Each node represents a single chunk of a document. Node color is determined by document type, making it easy to see at a glance which areas of the graph contain code, images, PDFs, and so on. Hover over a node to see:- The document name
- The document type and category
- The chunk index (position within the document)
- A preview of the chunk text (up to 200 characters)
Edges
Edges connect consecutive chunks within the same document. They show how a single document is distributed across the semantic space — a long document may span multiple clusters if it covers diverse topics.Clusters
When documents cluster together — their nodes appear close with no connecting edge between them — it means their content is semantically similar. This is how Prism surfaces unexpected thematic connections:- A research paper and a set of code files implementing the same algorithm will cluster together.
- Meeting notes and a project proposal discussing the same initiative will appear nearby.
- Multiple image descriptions depicting similar settings will group into a visual cluster.
Use cases
Discover connections
Find thematic relationships between documents you uploaded at different times — code files related to a PDF spec, or meeting notes referencing the same project as a contract.
Identify topic clusters
See which topics dominate your library by identifying the largest clusters and examining the document types they contain.
Spot outliers
Nodes that appear isolated far from any cluster represent documents with content unlike anything else in your library.
Validate ingestion
After uploading new documents, check whether they land near where you expect. If a technical document clusters with general notes rather than code files, consider reviewing how it was chunked.
The graph loads up to 100 vector chunks at a time. If your library has more than 100 chunks, the graph shows a representative sample. Upload fewer, larger documents to see a more complete picture in a single graph load.