Skip to main content
Semantic search lets you describe what you are looking for in plain language. Instead of matching exact words, Prism converts your query into a 768-dimensional vector and finds document chunks whose meaning is closest to your intent.
Keyword searchSemantic search
MatchingExact word or phrase matchMeaning and concept similarity
Query styleTerms and operatorsNatural language questions or descriptions
Finds synonymsNoYes
Understands contextNoYes
Examplemachine learning pythonhow do I train a neural network in Python?
Semantic search is especially useful when you cannot remember the exact wording of a document, or when you want to find content that discusses the same topic using different terminology.

​
Using the Search view

1

Open Search

Click Search in the left sidebar to open the Search view.
2

Type your query

Enter a question, phrase, or description in plain language. There is no need for special operators.
3

Apply filters (optional)

Use the document type filter to narrow results to a specific file format — for example, limit results to PDF files only, or to a specific code language.
4

Review results

Results are grouped by document. Each result shows the source document name, document type, category, the matching chunk of text, and a similarity score.

​
Understanding results

Each search result includes:
  • Document name — the file the matching chunk came from
  • Document type — file format (PDF, TS, PY, etc.)
  • Category — the auto-detected category (Code, Report, Legal, etc.)
  • Chunk text — the exact passage from the document that matched
  • Similarity score — a value between 0 and 1 indicating how closely the chunk meaning matches your query
Results are sorted by the highest similarity score per document.

​
Score threshold

The Search view applies a default score threshold of 0.5. Chunks with a similarity score below 0.5 are excluded from results. This keeps results relevant and reduces noise.
If you get no results, try rephrasing your query with more context. A specific question like “what are the payment terms in the contract?” typically returns better results than a single word like “payment”.

​
Filtering by document type

You can restrict search to one or more document types. This is useful when your library contains many different formats and you already know the file type of what you are looking for.
Select JS, TS, or PY from the document type filter before running your query. Prism will only search vector chunks from files with those extensions.
Select PDF from the document type filter. This is useful when searching a library that contains both PDFs and image files on the same topic.

​
How it works

Every search query goes through the same embedding pipeline used during document upload:
  1. Your query text is sent to Gemini’s text-embedding-004 model.
  2. Gemini returns a 768-dimensional vector representing the semantic meaning of the query.
  3. Qdrant performs a cosine similarity search against all indexed chunks for your account.
  4. Chunks above the score threshold are returned, sorted by relevance.
Because both documents and queries are embedded by the same model, meaning is preserved even when the exact wording differs.
Search results are always scoped to your account. You can only find documents you have uploaded.

​
Tips for better searches

  • Ask complete questions — “What does the license agreement say about sublicensing?” works better than “sublicense”.
  • Use domain language — technical terms specific to your documents produce more precise matches.
  • Try multiple phrasings — if results are weak, rephrase the question from a different angle.
  • Use type filters — if you know the file type, filtering narrows the search space and improves precision.
  • Check the chunk text — the matching passage shows exactly what Prism found. If the snippet is relevant but incomplete, open the source document for the full context.

Build docs developers (and LLMs) love

Get started for freeTalk to us