v0.2.0-alpha — gaps closed

Closes the three honest deferrals from v0.1.0-alpha. Every line item in the locked v0.1 plan §17 is now either shipped or has an explicit v0.3 home.

Install

go install github.com/aomerk/keeba/cmd/keeba@v0.2.0-alpha
keeba --version   # → keeba v0.2.0-alpha

What's new since v0.1.0-alpha

LLM-driven bench

ANTHROPIC_API_KEY=… keeba bench --llm anthropic --raw ../your-corpus

Each question is answered twice — wiki context (BM25 top-k) vs raw context (truncated dump) — by Claude. Token counts come from the API response (not estimates); confidence is self-rated 1-5. Output markdown includes side-by-side answers per question.

Inert without a key. Byte-count bench remains the no-key default.

Vector search

VOYAGE_API_KEY=… keeba index
keeba search --vector "JWT auth flow"

Embedder providers: Voyage AI, OpenAI. local is deferred to v0.3 (cybertron + ONNX MiniLM under the same interface).

Real llm.c worked example

Captured recipe + actual bench output checked into examples/llm-c/:

keeba: 2465.9× cheaper, 80.1× faster (5 questions; byte-count mode)

Replaces the placeholder demo-wiki number in the README.

Tests

~115 cases across 9 packages. golangci-lint v2.4.0 clean. Dogfood exercises every command (including the missing-key error paths) end-to-end.

What's v0.3

• Local sentence-transformers provider (cybertron + ONNX MiniLM).
• Direct execution of ingest agents (vs writing prompt templates).
• Cursor + Codex tool-config scaffolding alongside .mcp.json.
• Homebrew tap.
• 30-second screencast.

🤖 Released with Claude Code