Skip to main content
Smart recommendations surface documents in your library that are semantically similar to one you are already looking at. Instead of relying on tags, folders, or metadata, Prism compares the actual meaning of document content using their stored vector embeddings.

​
How to view recommendations

1

Open the Library

Navigate to Library in the left sidebar.
2

Select a document

Click on any document card. A detail panel opens showing the document’s information.
3

View related documents

The Recommended section in the detail panel lists documents that Prism has identified as semantically related to the selected file.

​
What each recommendation shows

For every recommended document, Prism displays:
FieldDescription
Document nameThe filename of the related document
Document typeFile format (PDF, TS, PY, IMAGE, etc.)
CategoryAuto-detected category (Code, Report, Legal, etc.)
Similarity scoreA value from 0.5 to 1.0 indicating how closely the document matches
Matching chunksUp to 3 representative text passages from the recommended document that were most similar to the selected document
Recommendations are ordered by similarity score from highest to lowest.

​
Similarity threshold

Only documents with a similarity score of 0.5 or higher appear as recommendations. This ensures that suggestions are genuinely related rather than superficially connected. Documents below this threshold are filtered out.

​
Cross-modal recommendations

Recommendations work across all file types. Prism compares meaning regardless of format:
  • A Python file implementing a machine learning model can recommend a PDF paper describing the same algorithm.
  • A meeting notes document can recommend a project proposal that discusses the same initiative.
  • An image of a UI wireframe can recommend a code file containing the corresponding component.
This cross-modal matching is possible because all content — whether extracted from text, code, or Gemini Vision image descriptions — goes through the same embedding model. The shared vector space enables like-meaning comparisons across unlike formats.
Recommendations are most useful when your library has at least 5–10 documents. With more documents, Prism has a larger pool to compare against and can surface more meaningful connections.

​
How recommendations are computed

When you select a document, Prism:
  1. Fetches a representative embedding vector from one of the document’s stored chunks.
  2. Searches Qdrant for the most similar chunks across your entire library, excluding chunks from the selected document itself.
  3. Groups matching chunks by their parent document and takes the highest-scoring chunk per document as that document’s similarity score.
  4. Keeps up to 3 matching chunk texts per recommended document — the three passages most similar to the selected document.
  5. Filters out any document whose best score is below 0.5.
  6. Returns the top 5 results sorted by score.
Recommendations are re-computed each time you select a document. They are not cached, so newly uploaded documents appear as recommendations immediately after their vectors are indexed.

​
Use cases

Select a source code file. If your library contains PDFs or Markdown files that document the same system, they will appear in recommendations alongside similar code files.
Select a contract or legal document. Recommendations may surface related agreements, amendments, or correspondence that reference the same parties or subject matter.
Select an image. Recommendations will surface other images with visually or thematically similar Gemini-generated descriptions, as well as any text documents that describe the same subjects.

Build docs developers (and LLMs) love

Get started for freeTalk to us