search_conversations

What It Does

search_conversations searches through indexed AI coding session histories using BM25F ranking with optional semantic search. It turns ephemeral conversation context into searchable institutional knowledge.

Every Claude Code session produces a JSONL file containing the full dialogue: questions asked, code discussed, errors encountered, solutions found. Without indexing, that knowledge disappears the moment the session ends. search_conversations makes it retrievable.

No Native Equivalent

There is no grep-based workflow that achieves this. The raw JSONL files contain tool_result dumps, base64 images, and multi-thousand-line tool outputs mixed in with the actual conversation. Grep against these files returns overwhelmingly noisy results. search_conversations strips this noise during indexing — tool results are truncated, images are replaced with [image] placeholders, and compact summaries are skipped in favor of the actual turn-pair content.

What Gets Indexed

Each conversation is decomposed into turn-pairs: a user question and the assistant’s response. The indexer extracts:

• User question — the actual question or instruction, stripped of system context
• Assistant answer — the substantive response, with tool outputs truncated
• Session metadata — session ID, timestamp, project path
• Conversation summaries — the final summary of a compacted session is indexed as a conversation_summary type

Tool use blocks are included but truncated to prevent grep-style noise. The full conversation is searchable by concept, not just by keyword.

When to Use

“We solved this before.” You encounter an EMFILE error or a Prisma migration failure. Before spending 30 minutes debugging, search your conversation history:

search_conversations("EMFILE error fix")

If a past session solved this exact problem, you get the solution in seconds.

“Why is the code like this?” You encounter a non-obvious pattern in the codebase — a manual SQL query where Prisma would work, or a seemingly unnecessary abstraction layer. The answer is often in a past conversation where the decision was made:

search_conversations("why manual SQL instead of Prisma for reports")

“Continue from yesterday.” A session was interrupted or compacted. Instead of re-reading the entire codebase to regain context:

search_conversations("auth refactor progress")

Onboarding. New to a project? Past conversations contain design decisions, rejected approaches, and architectural rationale that never made it into documentation:

search_conversations("architecture decisions", project="local/my-project")

Integration with codebase_retrieval

Conversation search is available as a query type in codebase_retrieval:

{
  "type": "conversation",
  "query": "why did we choose Redis for caching"
}

This lets you batch conversation search alongside code search in a single call — find the code and the discussion about it at the same time.

Key Parameters

• query (required) — natural language search query
• project — filter to a specific project. Defaults to the current project.
• limit — max results. Default 10.

Auto-Indexing

Conversations are indexed automatically. When CodeSift starts, it indexes conversations for the current project. A session-end hook re-indexes after each session completes, so new conversations are searchable in the next session without manual intervention.