Changelog

All notable changes to giddyanne will be documented in this file.

[1.5.0] - 2026-02-22

Added

• Shared embedding server: A global singleton service (embed_server.py) that loads the embedding model once and serves all project servers over a Unix socket. Project servers that previously loaded their own ~900MB model copy now proxy through the shared server at ~180MB each. Auto-starts on giddy up if not already running. Manage directly with giddy embed start|stop|status. Configurable via ~/.config/giddyanne/config.yaml (socket path, model, auto_start). Falls back to in-process embedding if the shared server is unavailable.
• giddy embed CLI command: giddy embed start, giddy embed stop, giddy embed status for managing the shared embedding server independently of any project.
• Global config (~/.config/giddyanne/config.yaml): Optional config for the shared embedding service — socket path, PID/log paths, model name, and auto_start toggle.
• Cross-encoder reranker (C4): Optional second-stage reranker using cross-encoder/ms-marco-MiniLM-L6-v2 (22.7M params). Over-fetches candidates (5x instead of 3x), runs existing pipeline (RRF, category bias, dedup), then re-scores with the cross-encoder before returning. Opt-in via reranker_model in .giddyanne.yaml. Uses CrossEncoder from sentence-transformers (no new dependency). Benchmarks: giddyanne recall 35%→80% (hybrid), MRR 0.34→0.90. test_2 FTS recall 60%→70%, MRR 0.40→0.70. test_1 MRR improved (+0.05-0.15) but semantic recall regressed 87%→67% on one query where the right file wasn't in the candidate pool. Disabled by default — best for repos where relevant files are retrieved but poorly ranked.
• --rerank CLI flag: Enable the cross-encoder reranker without a config file. giddy up --rerank and giddy bounce --rerank apply the default reranker model. For MCP, set GIDDY_RERANK=1 environment variable.
• Token-aware chunk splitting: After language-aware chunking, oversized chunks are recursively halved until they fit the model's token budget (256 tokens for MiniLM). Eliminates silent embedding truncation — content coverage went from 56% to 100%. Chunk count roughly doubles (313→740 for giddyanne repo). Benchmarks: test_2 recall 70%→77%, MRR 0.70→0.85. Semantic-only search now matches old hybrid baseline. Index times ~2x from more chunks. Ollama backend now impractical without batching (timed out at 600s on 32 files due to per-request HTTP overhead × ~2200 embed calls).
• Ollama embedding backend: GPU-accelerated embeddings via a local Ollama server. Opt-in via --ollama CLI flag, GIDDY_OLLAMA=1 env var, or ollama: true in config. Configurable URL and model (ollama_url, ollama_model). Default model: all-minilm:l6-v2.
• Truncation stats: /stats endpoint and giddy stats now report embedding coverage — how many chunks were truncated and what percentage of content is fully covered.

Changed

• BM25 field boosting (C3): Added fts_content column that repeats file path and symbol name 3x before raw content, simulating BM25F field-level boosting. FTS now indexes this enriched column instead of raw content. Benchmarks: test_1 full-text recall 63%→87%, "grid" query ranks grid.tsx #1. test_2 Stripe query ranks stripe.ts #1. giddyanne hybrid MRR holds at 1.00. Requires giddy clean --force && giddy up to rebuild index.
• Chunk context enrichment (C1): Prepend file path and function/class name to chunk content before embedding. Stored content unchanged (search results still show raw code). Benchmarks: giddyanne 75% recall (steady, "configuration and settings" query improved 0%→50%), test_2 80%→87%, test_1 77%→70% (one already-weak query regressed).
• Custom RRF with tunable weights (C2): Replaced LanceDB's black-box RRFReranker with our own weighted Reciprocal Rank Fusion. Runs semantic and FTS searches independently, merges ranked lists with configurable weights (semantic 1.2, FTS 0.5). Removes lancedb.rerankers dependency from hybrid search. Benchmarks: recall matches semantic-only (giddyanne 75%, test_2 70%, test_1 83%), MRR improved (giddyanne 0.65→1.00, test_2 0.65→0.70).

Added

• Quality benchmark script (benchmarks/quality.py): Reusable benchmark for measuring search recall, precision, and MRR across projects and search modes. Supports --mode all for side-by-side comparison of semantic, full-text, and hybrid search.

[1.4.1] - 2026-02-21

Fixed

• Version references: Updated all version strings across docs, Emacs package, and VSCode extension

[1.4.0] - 2026-02-21

Added

• Search result deduplication: Results are now deduplicated by file — only the best-scoring chunk per file is kept, freeing result slots for diverse files. Applied to all three search methods (semantic, full-text, hybrid). Over-fetches from LanceDB to compensate for collapsed results.
• File category scoring bias: Search scores are now weighted by file type — source code files rank at full weight (1.0×), test files at 0.8×, and docs at 0.6×. When a test file and its source have similar relevance, the source file surfaces first.

Changed

• Tree-sitter AST chunking: Replaced string-based separator splitting with tree-sitter AST parsing for code chunking. Functions, classes, and other definitions are now split at real node boundaries instead of pattern-matching on strings like "\ndef ". Falls back to blank-line splitting for languages without a tree-sitter grammar. Adds tree-sitter-language-pack dependency.
• Configurable embedding model: local_model in .giddyanne.yaml lets you swap embedding models. Each model gets its own index directory. Default remains all-MiniLM-L6-v2 — benchmarking showed nomic-ai/CodeRankEmbed (768d, code-specialized) indexes 5-6x slower and can't finish indexing a modest ~300-file repo on a MacBook Pro, with no quality gain to justify it.
• Cosine distance for vector search: Switched from L2 to cosine distance. Scores now display as meaningful 0-1 values instead of near-zero floats.
• Query prefix support: Models that require instruction prefixes on search queries (like CodeRankEmbed) are handled automatically. Document embeddings are unaffected.
• giddy install reads model from config: No longer hardcodes the model name — picks up local_model from .giddyanne.yaml or uses the default.

[1.3.2] - 2026-02-20

Added

• MCP status/health/stats tools: AI assistants can now check indexing progress, server health, and index statistics directly via MCP

Fixed

• giddy find waits for indexing on cold start: Previously returned a 503 error and exited if the server was still indexing. Now silently waits until ready, then returns results. Use --verbose to see progress.

[1.3.1] - 2026-02-20

Added

• Sitemap command in Emacs and VSCode: giddyanne-sitemap (Emacs) and Giddyanne: Sitemap (VSCode) show indexed files as a tree

Fixed

• Fix empty tree node for standalone file path groups in MCP sitemap output
• Fix unused variable warning in MCP server entrypoint
• Fix fmt.Println redundant newline warning in Go CLI

[1.3.0] - 2026-02-20

Added

• Sitemap command: new /sitemap endpoint, MCP sitemap tool, and giddy sitemap CLI command. Shows all indexed files, optionally with chunk counts and modification times (--verbose). API response includes a paths field with path and description for each configured group.
• File-tree sitemap output: sitemap displays files under their path groups with ├──/└── tree connectors, intermediate directories reconstructed. Verbose mode appends chunk count and mtime inline.
• Sitemap API tests: test coverage for /sitemap endpoint with and without config, including verbose mode and empty index

[1.2.1] - 2026-02-12

Added

• giddy install command: Downloads the embedding model (~90MB) upfront so giddy up starts immediately without a silent download pause

[1.2.0] - 2026-02-12

Added

• --version / -V flag for CLI (giddy --version prints giddy 1.2.0)
• MCP version tool returns the installed giddyanne version
• Single version source of truth: pyproject.toml drives all version strings — Go CLI reads it at build time via ldflags, Python reads via importlib.metadata
• Zero-config git repo support: giddy now works in any git repo without a .giddyanne.yaml config file. Auto-detects common source directories (src/, lib/, app/, tests/, docs/) and stores indexes in the system temp directory. Config file still takes priority when present.

[1.1.1] - 2025-12-31

Installation improvements.

Added

• Pre-built binaries for macOS (arm64, amd64) and Linux (amd64) attached to GitHub releases
• Flexible install location: make install BIN_DIR=/usr/local/bin

[1.1.0] - 2025-12-31

Search quality improvements, Emacs integration, and CLI polish.

Added

Search Quality - Language-aware chunking: split code at function/class boundaries using language-specific separators - Hybrid search: BM25 + semantic with --semantic, --full-text, --hybrid (default) modes - Model-specific vector stores: database path includes embedding model name - Faster initial import: parallelized file I/O + batched encode() calls

CLI - Shell completions for bash, zsh, and fish - Health metrics: index size, query latency via /health endpoint and giddy health - Prefix shortcuts: giddy f for find, giddy st for status, etc. - Better startup experience: friendlier first-run, clearer server feedback

Emacs - Full integration package with Vertico support and nerd-icons - Commands: giddyanne-find, giddyanne-up/down, giddyanne-status, giddyanne-health, giddyanne-log - Results grouped by file with pulse highlight on jump - Installation docs for Doom Emacs

VSCode - Extension with QuickPick search UI - Commands: Find, Start/Stop Server, Status, Health, Stream Logs - Results grouped by file with pulse highlight on jump - Keybinding: Cmd+Shift+G / Ctrl+Shift+G - Build with make vscode, install via .vsix

Languages - Language support expanded from 6 to 28 languages: - Programming: C, C++, C#, Dart, Elixir, Java, Kotlin, Lua, Objective-C, PHP, R, Ruby, Scala, Shell, Swift, Zig - Markup & config: CSS, HTML, JSON, Markdown, TOML, YAML

Fixed

• INFO logs no longer appear without --verbose flag
• giddy health shows current state when server is down (like giddy status)
• giddy clean finds model-specific index paths
• --files-only --json returns only path and lines, not content

Breaking Changes

• Database path includes model name. Migration: mkdir -p .giddyanne/all-MiniLM-L6-v2 && mv .giddyanne/vectors.lance .giddyanne/all-MiniLM-L6-v2/
• Removed OpenRouter support. Remove GIDDY_OPENROUTER_API_KEY env var and openrouter_model from config

[1.0.0] - 2025-12-30

First stable release. Semantic code search for codebases via CLI and MCP server.

Features

Core - Semantic code search using local sentence-transformers embeddings - LanceDB vector store with chunk-based schema - Hybrid chunking: split on blank lines, merge small chunks, split large with overlap - File watcher for automatic re-indexing on changes - Search query caching with hit statistics - File timestamp tracking to skip unchanged files on startup - Index reconciliation removes stale files on startup - Glob patterns in ignore settings

CLI (giddy) - giddy search <query> - search with auto-start if server not running - giddy start - start the server daemon - giddy stop - stop the server - giddy status - show server status and indexing progress - giddy monitor - stream server logs - giddy clean - clear the search index - giddy init - generate config file prompt - giddy mcp - run MCP server for Claude Code - --verbose flag for debug logging - --port, --host flags for server configuration

HTTP API - POST /search - semantic search - GET /status - indexing progress - GET /health - liveness check - GET /stats - index statistics

MCP Server - Single search tool for Claude Code integration - Description embeddings for better context matching

Configuration - .giddyanne.yaml project config with paths and descriptions - Configurable chunking parameters