Retrace Docs
Features

Semantic Search

Search your conversations by meaning, not just keywords.

What it does

Semantic search finds conversations based on what they mean, not just the exact words used. A query like "feeling stuck at work" finds conversations about career frustration even if nobody said those exact words.

How it works

When you import conversations, Retrace breaks them into segments and generates an embedding for each one. An embedding is a mathematical representation of the meaning of a text. When you search, your query gets the same treatment, and Retrace finds segments whose embeddings are closest in meaning.

You can search in three modes:

  • Keyword: Traditional text matching. Finds exact words.
  • Semantic: Meaning-based matching. Finds related concepts.
  • Hybrid: Combines both for the best results.

What data it uses

  • Conversation segments (chunks of messages grouped by time and topic)
  • Embeddings generated locally on your machine

What it produces

Search results ranked by relevance, with similarity scores showing how close each result is to your query. Each result links to the specific segment within the conversation.

Privacy

  • Embeddings are generated locally using a small model running on your machine
  • No search queries or conversation data are sent to any external service
  • The embedding model and vector index are stored in your local SQLite database

On this page