Most retrieval stacks assume you must first turn files into embeddings and store them in a vector database. Sirchmunk bets the opposite: for a lot of real work, the raw text is already searchable, and paying the indexing tax up front is the actual bottleneck. It runs a ripgrep-based agentic search directly over your files and only spends LLM tokens when a candidate match genuinely needs reasoning — so "index time" drops to zero and cost scales with query difficulty, not corpus size.
What Sets It Apart
- No vector store, no re-indexing: drop files in and query immediately. When data changes, there is nothing to rebuild — the index that would go stale simply does not exist.
- LLM only when it earns its keep: cheap grep-style candidate gathering first, model calls reserved for evidence that needs judgment. A FAST mode does file-level dedup and score pruning; a DEEP mode adds Monte Carlo evidence sampling for 10–30s comprehensive answers.
- Self-evolving knowledge clusters: results are persisted as DuckDB/Parquet clusters with deterministic SHA256 IDs and reused via semantic similarity, so repeated questions get faster without a static snapshot.
- Agent-native surface: ships an MCP server (Claude Desktop, Cursor), an OpenClaw skill, an SSE-streaming Search API, a CLI, and a Next.js web UI.
Who It's For
Great fit if you want retrieval over a live, frequently changing local corpus, or you are wiring search into an agent and resent maintaining a vector-DB pipeline for it. The token-metered design rewards workloads where most queries are shallow and only a few need deep reasoning. Look elsewhere if you need semantic recall over millions of near-duplicate passages where dense embeddings genuinely win, or if your data is static enough that a one-time index is simply cheaper than per-query LLM calls.
