vera-ai 0.12.0

Bootstrap installer and wrapper for the Vera CLI

vera-ai

Code search for AI agents. Vera indexes your codebase using tree-sitter parsing and hybrid search (BM25 + vector similarity + cross-encoder reranking), then returns ranked code snippets as structured JSON.

This package downloads and wraps the native Vera binary for your platform. On musl-based Linux (Alpine, NixOS), the correct static binary is selected automatically. Set VERA_TARGET to override target detection (e.g., VERA_TARGET=x86_64-unknown-linux-musl uvx vera-ai install).

Current benchmark snapshot: on Vera's local 21-task, 4-repo release benchmark, v0.7.0 reaches 0.78 Recall@5, 0.83 Recall@10, 0.91 MRR@10, and 0.84 nDCG@10 with the local Jina CUDA ONNX stack. Full details live in the main repo docs.

Install

pip install vera-ai

vera-ai setup only configures the backend. Run vera-ai agent install to set up skill files for your agents (interactive by default, or pass --client and --scope for non-interactive use). The interactive flow can also update AGENTS.md / CLAUDE.md style project instructions for you.

Usage

# Optional: install skill files for your agents
vera-ai agent install

# Index a project
vera-ai index .

# Search
vera-ai search "authentication middleware"

# Local ONNX inference (no API keys needed. downloads models automatically)
vera-ai index . --onnx-jina-cpu
vera-ai search "error handling" --onnx-jina-cpu

# Optional local CodeRankEmbed preset
vera-ai setup --code-rank-embed --onnx-jina-cuda

# GPU acceleration (NVIDIA/AMD/DirectML/CoreML/OpenVINO)
vera-ai index . --onnx-jina-cuda

# Diagnose or repair local setup issues
vera-ai doctor
vera-ai doctor --probe
vera-ai repair
vera-ai upgrade

vera-ai doctor --probe runs a deeper read-only ONNX session check. vera-ai upgrade shows the binary update plan and can apply it when the install method is known.

On GPU backends, Vera uses a free-VRAM-aware batch ceiling and sequence-aware local micro-batching, and it reuses learned device-specific batch windows across runs.

What you get

• 60+ languages via tree-sitter AST parsing
• Hybrid search: BM25 keyword + vector similarity, fused with Reciprocal Rank Fusion
• Cross-encoder reranking for precision
• Markdown codeblock output by default with file paths, line ranges, and optional symbol info (use --json for compact JSON, --raw for verbose output, --timing for step durations)

For full documentation, including custom local ONNX embedding models and manual install steps, see the GitHub repo.