AI code search that skips embeddings entirely

What it does

Probe is a context engine for AI agents navigating large codebases. It combines tree-sitter AST parsing with Elasticsearch-style boolean queries to return entire functions, classes, or structs instead of line fragments or arbitrary text chunks. A built-in agent layer handles multi-step reasoning, extraction, and code editing through MCP or direct CLI integration.

The interesting bit

The project’s central bet is that AI agents don’t need embedding search at all. Since the LLM can translate natural language into precise boolean queries like verify_credentials OR authenticate, Probe skips vector databases, embedding APIs, and indexing overhead entirely. It relies on deterministic BM25/TF-IDF ranking with SIMD acceleration and session deduplication to keep context windows clean.

Key highlights

• Returns entire semantic blocks via tree-sitter parsing instead of splitting mid-function
• Zero setup: no indexing, no embedding models, no external API calls for search; runs fully offline
• Deterministic results: same query always returns the same output, with no model variance or stale indexes
• Token-aware budgets and session deduplication to avoid repeating context across agent loops
• Multi-provider built-in agent with MCP, CLI, and Node.js SDK integrations

Verdict

Worth a look if you’re building AI coding tools or want deterministic, offline code search without managing vector databases. Skip it if you’re already happy with ripgrep and don’t need structured context blocks for LLM consumption.