Features
Knowledge Extraction
How Retrace breaks down conversations into searchable, browsable segments.
What it does
Long conversations are hard to search and browse. Knowledge extraction automatically breaks conversations into meaningful segments (chunks grouped by time and topic), then extracts a summary, topics, and sentiment for each one.
Instead of scrolling through thousands of messages, you can search "career doubts with Hugo" and land on the exact segment.
How it works
- Segmentation: Conversations are split into time-bounded chunks. A new segment starts when there's a significant gap in time or a topic shift.
- Summarization: A local LLM reads each segment and generates a short summary.
- Topic extraction: Key topics are identified and tagged (e.g., "career", "travel", "health").
- Sentiment analysis: Each segment gets a sentiment score (positive, neutral, negative).
- Embedding: A vector embedding is generated for semantic search.
What data it uses
- Raw messages from all imported conversations
- Timestamps and sender information
What it produces
- Segments with summaries, topics, and sentiment
- Browsable timelines grouped by week or month
- Searchable index for both keyword and semantic queries
Privacy
- Summarization and topic extraction run locally using a small LLM on your machine
- Embeddings are generated locally
- No conversation content is sent to any external service during extraction