Configuration

Giddyanne works in any git repo without configuration. A .giddyanne.yaml file is optional but recommended for better search quality.

Zero-Config Mode

In a git repo without .giddyanne.yaml, giddyanne automatically: - Uses the git root as the project root - Indexes common directories (src/, lib/, app/, tests/, docs/) or the whole project if none exist - Respects .gitignore patterns - Stores indexes in the system temp directory ($TMPDIR/giddyanne/<name>-<hash>/)

Indexes rebuild automatically if the temp directory gets cleaned. This is fine — giddy up is fast.

Config File

For better search quality, add a .giddyanne.yaml to your project root. Path descriptions are embedded alongside content, improving semantic matching.

Minimal Config

paths:
  - path: .
    description: My app

That's it. Only supported language files are indexed, and .gitignore is respected automatically.

Annotated Example

paths:
  - path: src/
    description: Core application source code

  - path: src/auth/
    description: Authentication, login, sessions, permissions

  - path: tests/
    description: Test suite and fixtures

Descriptions are embedded alongside content for better semantic matching. More specific descriptions improve search quality.

Generating a Config

Run giddy init to get a prompt you can paste into Claude or another LLM:

claude "follow the instructions: $(giddy init)"

The LLM will analyze your codebase and generate an appropriate config.

Full Example

All available options:

paths:
  # Index the entire project with a general description
  - path: .
    description: Web application with REST API

  # Override with more specific descriptions for key directories
  - path: src/api/
    description: REST endpoints, request handlers, middleware

  - path: src/auth/
    description: Authentication, JWT tokens, session management, permissions

  - path: src/db/
    description: Database models, migrations, query builders

  - path: src/utils/
    description: Shared utilities, helpers, validation functions

  - path: tests/
    description: Test suite, fixtures, mocks

settings:
  # File handling
  max_file_size: 1000000           # Skip files > 1MB
  ignore_patterns:                  # Beyond .gitignore
    - "*.generated.*"
    - "*.min.js"
    - "vendor/"

  # Server
  host: "0.0.0.0"                  # Bind address
  port: 8000                        # HTTP port

  # Chunking (defaults work well for most codebases)
  min_chunk_lines: 10
  max_chunk_lines: 50
  overlap_lines: 5

  # Embedding model
  local_model: "all-MiniLM-L6-v2"

  # Ollama backend (optional, disabled by default)
  ollama: true
  ollama_url: "http://localhost:11434"
  ollama_model: "all-minilm:l6-v2"

  # Reranker (optional, disabled by default)
  reranker_model: "cross-encoder/ms-marco-MiniLM-L6-v2"

Settings Reference

Paths

Field	Required	Description
path	Yes	File or directory to index (relative to config file)
description	No	Human-readable description, embedded alongside content for better search

File Handling

Setting	Default	Description
max_file_size	1000000	Skip files larger than this (bytes)
ignore_patterns	[]	Additional patterns to ignore (uses gitignore syntax)

Server

Setting	Default	Description
host	"0.0.0.0"	Network interface to bind to
port	8000	HTTP port for the server

Chunking

Code is split into chunks for indexing. These settings control chunk boundaries:

Setting	Default	Description
min_chunk_lines	10	Minimum lines per chunk
max_chunk_lines	50	Maximum lines per chunk
overlap_lines	5	Lines of overlap between chunks

Defaults work well for most codebases. Smaller chunks give more precise results but may lose context.

Embedding

Setting	Default	Description
local_model	"all-MiniLM-L6-v2"	Sentence-transformers model name

With a config file, the database is stored at .giddyanne/<model-name>/vectors.lance. In zero-config mode, it's stored in the system temp directory. Each model gets its own index, so you can switch models without losing data.

Available models: - all-MiniLM-L6-v2 (default) — Good balance of speed and quality - all-mpnet-base-v2 — Higher quality, slower - paraphrase-MiniLM-L3-v2 — Faster, lower quality - nomic-ai/CodeRankEmbed — Code-specialized (768d), but 5-6x slower indexing with no measurable quality gain. Unusable on laptop hardware for repos over ~100 files.

Ollama

Use a local Ollama server for GPU-accelerated embeddings instead of the default sentence-transformers (CPU).

Setting	Default	Description
ollama	false	Enable Ollama backend
ollama_url	"http://localhost:11434"	Ollama server URL
ollama_model	"all-minilm:l6-v2"	Ollama embedding model

settings:
  ollama: true
  ollama_model: "all-minilm:l6-v2"

Without a config file, use the --ollama CLI flag instead:

giddy up --ollama

For MCP, set the GIDDY_OLLAMA=1 environment variable in your MCP config.

Requires Ollama running locally with the model pulled (ollama pull all-minilm:l6-v2).

Reranker

Setting	Default	Description
reranker_model	"" (disabled)	Cross-encoder model for result reranking

The reranker is a second-stage scorer that re-evaluates the top search results using a cross-encoder. It improves ranking quality at the cost of slightly slower queries (the model scores each result individually).

settings:
  reranker_model: "cross-encoder/ms-marco-MiniLM-L6-v2"

Without a config file, use the --rerank CLI flag instead:

giddy up --rerank

For MCP, set the GIDDY_RERANK=1 environment variable in your MCP config.

Global Config

The shared embedding server has its own config at ~/.config/giddyanne/config.yaml. This is separate from per-project .giddyanne.yaml files.

# All fields are optional — defaults shown
embed_model: "all-MiniLM-L6-v2"
auto_start: true
socket_path: "~/.local/state/giddyanne/embed.sock"
pid_path: "~/.local/state/giddyanne/embed.pid"
log_path: "~/.local/state/giddyanne/embed.log"

Setting	Default	Description
embed_model	"all-MiniLM-L6-v2"	Model the shared server loads
auto_start	true	Auto-start the embed server when a project server starts
socket_path	~/.local/state/giddyanne/embed.sock	Unix socket for embed server
pid_path	~/.local/state/giddyanne/embed.pid	PID file location
log_path	~/.local/state/giddyanne/embed.log	Log file location

Set auto_start: false to disable the shared server and fall back to in-process embedding (each project loads its own model copy).