ls-conversation-lookup

To search Claude Code conversation history, you can use an ID prefix to resolve file paths or perform BM25 keyword search across indexed JSONL messages. It retrieves relevant conversations by ranking user and assistant transcripts stored in a local SQLite database.

Semantic vector search on conversation transcripts uses sentence-transformer embeddings, specifically all-MiniLM-L6-v2, to match query intent. It compares vector similarities against indexed messages to find contextually relevant results without relying on exact keyword matches.

Yes, hybrid retrieval requires both an FTS5 index and sentence-transformer embeddings. It combines BM25 keyword ranking with semantic vector similarity, meaning you need both dependencies installed to execute hybrid search queries across your conversations.

Yes, conversation lookup supports optional project and time filtering to narrow results. You can constrain searches across the 800K+ indexed messages to find conversations specific to certain projects or within defined timeframes.

The best way to summarize Claude Code JSONL artifacts is using the built-in summarization feature, which utilizes a local LLM via LM Studio. Alternatively, you can extract a condensed transcript without an LLM for faster review.

BM25 keyword search is better for conversation lookup when you need exact term matching within the FTS5 index. Semantic search provides contextual matching using embeddings, while hybrid retrieval combines both for comprehensive ranking accuracy.