ogrep 0.7.2

Semantic grep for codebases - local-first, SQLite-backed, with local or cloud embeddings

ogrep

Semantic grep for codebases — local-first, SQLite-backed, and built for Claude Code Skills (not MCP).

ogrep helps you search code by meaning, not just keywords. It builds a local semantic index (.ogrep/index.sqlite by default) and retrieves the most relevant code chunks for questions like:

• "where is authentication handled?"
• "how are API errors mapped to exceptions?"
• "where do we open DB connections and run queries?"
• "what kind of API key mechanism do we use?"

GitHub: github.com/gplv2/ogrep-marketplace Website: ogrep.be — quick overview

---

What's New in v0.7.2

Breaking Change (v0.7.2)

• JSON Output is Now Default — All commands output JSON by default for AI/machine use. Use --no-json for human-readable text output. The --json flag is still accepted for backwards compatibility.

Major Features (v0.7.0+)

• AST-Aware Chunking — Use --ast to chunk code by function/class/method boundaries instead of arbitrary line counts. This produces semantically coherent chunks that dramatically improve search accuracy for code-related queries.

• Cross-Encoder Reranking — Add --rerank to apply a cross-encoder model for high-precision ranking of search results. Solves the "right file is in top 30 but not #1" problem.

• RRF Hybrid Fusion — Reciprocal Rank Fusion replaces alpha weighting as the default hybrid search method. Combines results by position rather than raw scores for more robust ranking.

• AST Mode Tracking — Index now tracks whether AST chunking was used. Query results include hints when the index could benefit from AST mode.

Improvements

• Index Change History — Track what changed with ogrep log, useful for AI tool integration
• Fusion Method in JSON — Query stats now include fusion_method to show which method was used
• Graceful Degradation — Works reliably with or without optional features (reranking, FTS5, GPU, AST)

Recent (v0.6.x)

• Cross-file chunk deduplication (up to 80% embedding cost savings)
• Relative confidence scoring (compares to top result, not fixed thresholds)
• Graceful Ctrl-C handling with recovery messages

---

Installation

Option A: pip (recommended)

pip install ogrep

Option B: pipx (isolated environment)

pipx install ogrep

Note: pipx sometimes has issues. If you encounter problems, use pip instead.

Option C: Claude Code Marketplace + Plugin

# Add the marketplace
/plugin marketplace add gplv2/ogrep-marketplace

# Install the plugin
/plugin install ogrep@ogrep-marketplace

It will ask where to install. Use 'user' mode — local mode can cause path issues when working on multiple codebases.

Optional Extras

# AST-aware chunking (recommended for code search)
pip install "ogrep[ast]"           # Python/JS/TS/Go/Rust support
pip install "ogrep[ast-all]"       # All 13 supported languages

# Cross-encoder reranking (high-precision ranking)
pip install "ogrep[rerank]"        # sentence-transformers

# Other extras
pip install "ogrep[speed]"         # Faster scoring with numpy
pip install "ogrep[mcp]"           # MCP server support

# Combine extras
pip install "ogrep[ast,rerank]"    # AST + reranking

---

Quick Start

With OpenAI

export OPENAI_API_KEY="sk-..."

ogrep index .                              # Index current directory
ogrep query "where is auth handled?" -n 10 # Semantic search
ogrep status                               # Check index stats

With LM Studio (Local, Free)

# 1. Install LM Studio from https://lmstudio.ai
# 2. Download and load a model
lms get nomic-embed-text-v1.5 -y
lms load nomic-ai/nomic-embed-text-v1.5-GGUF -y
lms server start

# 3. Point ogrep to local server
export OGREP_BASE_URL=http://localhost:1234/v1

# 4. Index and query
ogrep index . -m nomic
ogrep query "database connection handling" -m nomic

See LOCAL_EMBEDDINGS_GUIDE.md for detailed setup and tuning.

---

AST-Aware Chunking

The biggest accuracy improvement for code search. Instead of splitting by arbitrary line counts, AST chunking respects function, class, and method boundaries.

Why AST Chunking Matters

Without AST (line-based chunks):

Lines 55-115 (one chunk):
  - End of ClassA
  - Start of ClassB  ← Semantic mixing!
  - Beginning of method foo()

With AST chunking:

Chunk 1: ClassA (complete)
Chunk 2: ClassB.foo() method
Chunk 3: ClassB.bar() method

Usage

# Install AST support
pip install "ogrep[ast]"           # Python/JS/TS/Go/Rust
pip install "ogrep[ast-all]"       # All 13 languages

# Index with AST chunking
ogrep index . --ast

# Check if index uses AST
ogrep status
# Output: AST Mode: enabled

# Reindex existing index with AST
ogrep reindex . --ast

Supported Languages

Language	Extension	Package
Python	.py	ogrep[ast]
JavaScript	.js	ogrep[ast]
TypeScript	.ts, .tsx	ogrep[ast]
Go	.go	ogrep[ast]
Rust	.rs	ogrep[ast]
C	.c, .h	ogrep[ast-all]
C++	.cpp, .hpp	ogrep[ast-all]
Java	.java	ogrep[ast-all]
Ruby	.rb	ogrep[ast-all]
PHP	.php	ogrep[ast-all]
C#	.cs	ogrep[ast-all]
Scala	.scala	ogrep[ast-all]
Kotlin	.kt	ogrep[ast-all]

Files in unsupported languages fall back to line-based chunking automatically.

---

Cross-Encoder Reranking

Solves the "right file in top 30 but not #1" problem. Cross-encoders process (query, document) pairs together, providing much higher precision than bi-encoder embeddings.

How It Works

Query → Stage 1: Fast Retrieval (embeddings + BM25) → Top 50 candidates
                              ↓
      Stage 2: Slow Reranking (cross-encoder) → Top 10 results

The cross-encoder sees both query AND document together, so it can model fine-grained relationships that embeddings miss.

Usage

# Install reranking support
pip install "ogrep[rerank]"

# Enable reranking (fetches 50, reranks, returns top -n)
ogrep query "where is authentication?" -n 10 --rerank

# Custom rerank pool size
ogrep query "where is authentication?" -n 10 --rerank-top 100

Reranking Model

ogrep uses BAAI/bge-reranker-v2-m3 by default (~300MB, auto-downloaded on first use). This model works well with code and is multilingual.

Configure via environment:

export OGREP_RERANK_MODEL=BAAI/bge-reranker-v2-m3
export OGREP_RERANK_TOPN=50

---

Search Modes & Hybrid Fusion

ogrep supports three search modes via --mode (or -M):

Mode	Best For	How It Works
hybrid	General use (default)	RRF fusion of semantic + keyword
semantic	Conceptual questions	Embeddings only — "where is auth handled?"
fulltext	Exact identifiers	FTS5 keywords — "def validate_token"

# Default: hybrid (best of both worlds)
ogrep query "user authentication" -n 10

# Pure semantic (meaning-based)
ogrep query "how are errors handled" --mode semantic

# Pure keyword (exact matches)
ogrep query "class AuthMiddleware" --mode fulltext

RRF Fusion (Default)

Reciprocal Rank Fusion combines results by position, not raw scores:

rrf_score = 1/(k + semantic_rank) + 1/(k + fulltext_rank)

Benefits:

• No tuning required (k=60 is standard)
• Handles score distribution differences
• Results appearing in both lists are properly boosted

Legacy Alpha Weighting

If you prefer the old score-based fusion:

export OGREP_FUSION_METHOD=alpha
export OGREP_HYBRID_ALPHA=0.7  # 70% semantic, 30% keyword

---

AI Tool Integration

All commands output JSON by default — optimized for AI tools, scripts, and programmatic contexts. Use --no-json for human-readable text output.

JSON Output (Default)

ogrep query "database connections"

{
  "query": "database connections",
  "results": [
    {
      "rank": 1,
      "chunk_ref": "src/db.py:2",
      "path": "/home/user/project/src/db.py",
      "relative_path": "src/db.py",
      "start_line": 45,
      "end_line": 78,
      "score": 0.8923,
      "confidence": "high",
      "language": "python",
      "text": "def connect_to_database(config):\n    ..."
    }
  ],
  "stats": {
    "total_results": 10,
    "total_chunks": 234,
    "search_time_ms": 45,
    "search_mode": "hybrid",
    "fusion_method": "rrf",
    "reranked": false,
    "fts_available": true,
    "index_model": "text-embedding-3-small",
    "index_dimensions": 1536,
    "ast_mode": true,
    "confidence_summary": {"high": 3, "medium": 5, "low": 2}
  }
}

AST Mode Hints

When querying an index built without AST chunking, JSON output includes a hint:

{
  "results": [...],
  "stats": { "ast_mode": false },
  "hint": "Index was built without AST chunking. For better semantic boundaries, run: ogrep reindex . --ast"
}

Status Check

ogrep status

{
  "database": ".ogrep/index.sqlite",
  "status": "indexed",
  "indexed": true,
  "files": 45,
  "chunks": 234,
  "model": "text-embedding-3-small",
  "dimensions": 1536,
  "ast_mode": true,
  "size_bytes": 2456789,
  "size_human": "2.3 MB"
}

For Claude Code Skills

JSON output is now the default. The Claude Code Skill should:

• Use --refresh to ensure results reflect current codebase state
• Check stats.ast_mode and suggest ogrep reindex . --ast if false
• Use --no-json only if human-readable output is needed

---

CLI Commands

All commands output JSON by default. Use --no-json for human-readable text.

Command	Description
ogrep index .	Index current directory
ogrep index . --ast	Index with AST-aware chunking
ogrep index . --list	Preview files before indexing
ogrep query "text" -n 10	Search (hybrid mode by default)
ogrep query "text" --rerank	Search with cross-encoder reranking
ogrep query "text" --no-json	Human-readable output
ogrep query "text" --mode semantic	Pure semantic search
ogrep query "text" --mode fulltext	Keyword search (FTS5)
ogrep chunk "path:N" -C 1	Get chunk with context
ogrep status	Show index statistics
ogrep device	Check GPU/CPU for reranking
ogrep health	Full database diagnostics
ogrep health --vacuum	Reclaim space and defragment
ogrep health --full	Vacuum + rebuild FTS5 + integrity check
ogrep log	Show index change history
ogrep delete "path"	Remove files from index
ogrep reset -f	Delete index
ogrep reindex . --ast	Rebuild with AST chunking
ogrep clean --vacuum	Remove stale entries
ogrep models	List available embedding models
ogrep tune .	Auto-tune chunk size
ogrep benchmark .	Compare all models

---

Real-world Scenarios

1) Rebuilding legacy systems by behavior (my primary use)

When you inherit a legacy codebase (PHP spaghetti, mixed triggers/procs, half-documented business logic), "fixing in place" often becomes a trap: every change risks regressions, and understanding intent takes forever.

ogrep supports a different approach:

• Understand intent → extract behavior → rebuild cleanly
• Identify what the system does (invoices, device provisioning, auth, state transitions, edge cases)
• Reconstruct a behavioral spec and implement a new, maintainable system that mimics the original outcomes — without dragging the old architecture along.

Think "software archaeology": you're not searching for a string, you're searching for meaning.

2) Turning "token blackholes" into a cheap retrieval step

The common workflow is painful and expensive:

grep → copy/paste huge files → LLM reads everything → repeat → burn tokens

ogrep flips that:

• You index once (embeddings stored in SQLite)
• Queries retrieve top-K relevant snippets fast
• You only send the small, relevant results to an LLM when needed

Validate the claim: ogrep itself does not need a chat LLM to work. It uses embeddings for indexing + query retrieval.

• With local embeddings (LM Studio), embedding cost is effectively free
• With OpenAI embeddings, you still pay embedding tokens during indexing (and a tiny amount per query), but you avoid the "paste the repo into a chat model" cost explosion

3) Fast navigation through unknown repos

• Find where a feature "really" lives (even if naming is inconsistent)
• Trace flows like "request → validation → persistence → side effects"
• Discover the real entry points, glue code, and hidden coupling

4) Safer refactors and migrations

• Locate the real "source of truth" logic before rewriting
• Identify duplicated or divergent implementations
• Build a migration plan based on actual code paths, not guesswork

---

Embedding Providers

Choose your embedding source:

Provider	Cost	Privacy	Setup
OpenAI API	$0.02/M tokens	Cloud	Just add OPENAI_API_KEY
LM Studio (local)	Free	100% local	Run lms server start

Setting up environment

# OpenAI (cloud)
export OPENAI_API_KEY="sk-..."
ogrep index . -m small

# LM Studio (local, free, offline)
export OGREP_BASE_URL=http://localhost:1234/v1
ogrep index . -m nomic

Using direnv for autoloading .env (optional)

Install direnv and add to your .bashrc:

eval "$(direnv hook bash)"

Create a .envrc file in the base dir:

# Auto-load .env when entering directory
dotenv

Allow it:

direnv allow

---

Confidence Scores

Results include confidence levels to help you decide how much to trust them:

Confidence	Score	Guidance
high	0.85+	Trust and use directly
medium	0.70-0.84	Use but verify context
low	0.50-0.69	Consider alternative queries
very_low	<0.50	Likely not relevant

Tuning Confidence Thresholds

The default thresholds work well for well-documented codebases. For legacy code with sparse comments:

export OGREP_CONFIDENCE_HIGH=0.60
export OGREP_CONFIDENCE_MEDIUM=0.45
export OGREP_CONFIDENCE_LOW=0.35

Understanding Low Scores

Semantic search works best when code has good comments, docstrings, or descriptive variable names. Dense implementation code with few comments tends to score lower.

If you're getting consistently low scores:

1. Use AST chunking — ogrep reindex . --ast for better semantic boundaries
2. Try reranking — --rerank for more accurate ordering
3. Try code-like queries — match the terminology in the code
4. Use fulltext mode — for exact identifiers: --mode fulltext
5. Lower thresholds — for legacy codebases (see above)
6. Check chunk context — use ogrep chunk "path:N" -C 2 to expand

---

Chunk Navigation

Found something interesting? Expand the context:

# Get chunk by reference (from query results)
ogrep chunk "src/auth.py:2"

# Include surrounding chunks
ogrep chunk "src/auth.py:2" --before 1    # 1 chunk before
ogrep chunk "src/auth.py:2" --after 1     # 1 chunk after
ogrep chunk "src/auth.py:2" --context 1   # 1 before AND after

---

Embedding Models

OpenAI Models (Cloud)

Model	Alias	Dimensions	Price	Best For
text-embedding-3-small	small	1536	$0.02/M	Most use cases (default)
text-embedding-3-large	large	3072	$0.13/M	High-accuracy, multi-language
text-embedding-ada-002	ada	1536	$0.10/M	Legacy compatibility

Local Models (via LM Studio)

Model	Alias	Dimensions	Accuracy	Notes
all-MiniLM-L6-v2	minilm	384	96%	Best accuracy, smallest (~25MB)
nomic-embed-text-v1.5	nomic	768	72%	Large context window (8192 tokens)
bge-base-en-v1.5	bge	768	52%	Fallback option
bge-m3	bge-m3	1024	TBD	Multi-lingual (100+ languages)

Important: Query model must match index model. Use ogrep status to check.

---

Smart Defaults

ogrep is optimized for source code search out of the box.

Source-Only Indexing

By default, ogrep indexes only source files and excludes:

Category	Examples
Docs	*.md, *.txt, *.rst, docs/*
Config	*.json, *.yaml, *.toml, .editorconfig
Secrets	.env, secrets.*, credentials.*
Build	dist/*, build/*, *.min.js
Binary	Images, fonts, media, archives
Databases	*.sqlite, *.db, *.sql, *.dump
Data files	*.csv, *.tsv, *.xml, *.dat
Backups	*.old, *.bak, *.backup, *.orig, *~
Temp files	*.tmp, *.temp, *.swp
Lock files	package-lock.json, yarn.lock, poetry.lock

Skipped directories: .git/, .svn/, .hg/, node_modules/, .venv/, __pycache__/, .ogrep/

Smart Embedding Reuse

ogrep minimizes API costs with intelligent incremental indexing:

$ ogrep index .
Indexed into .ogrep/index.sqlite
  Files: 3 indexed, 42 skipped
  Chunks: 12 total (9 reused, ~900 tokens saved)

Edit Pattern	Without Reuse	With Reuse	Savings
Edit 1 line in 300-line file	5 embeds	1 embed	80%
Append function to file	5 embeds	1 embed	80%
No changes	5 embeds	0 embeds	100%

---

File Filtering

Include Normally-Excluded Files

ogrep index . -i '*.md'             # Include markdown
ogrep index . -i '*.md' -i '*.json' # Multiple patterns

Add Extra Exclusions

ogrep index . -e 'test_*' -e '*_test.py'  # Exclude tests
ogrep index . -e 'fixtures/*'              # Exclude directories

.ogrepignore File

Create a .ogrepignore file for permanent exclusions:

# .ogrepignore - glob patterns like .gitignore
*.sql
*.dump
migrations/*
legacy/*

---

Auto-Tuning

Different models and codebases have different optimal chunk sizes:

ogrep tune . -m nomic

Testing chunk size 30... accuracy=0.72 (5/5 hits)  <-- OPTIMAL
Testing chunk size 45... accuracy=0.56 (4/5 hits)
Testing chunk size 60... accuracy=0.36 (3/5 hits)

Recommended chunk size: 30 lines

Save & Apply

ogrep tune . -m nomic --save        # Save to .env
ogrep tune . -m nomic --apply       # Reindex immediately
ogrep tune . -m nomic --save --apply # Both

---

Environment Variables

Variable	Description	Default
OPENAI_API_KEY	OpenAI API key (required for cloud)	—
OGREP_BASE_URL	Local server URL (e.g., LM Studio)	—
OGREP_MODEL	Default embedding model	Smart default*
OGREP_CHUNK_LINES	Tuned chunk size	Model default
OGREP_DIMENSIONS	Embedding dimensions	Model default
OGREP_SEARCH_MODE	Default search mode	hybrid
OGREP_FUSION_METHOD	Hybrid fusion method	rrf
OGREP_HYBRID_ALPHA	Semantic weight (if using alpha)	0.7
OGREP_RERANK_MODEL	Cross-encoder model	BAAI/bge-reranker-v2-m3
OGREP_RERANK_TOPN	Candidates to rerank	50
OGREP_CONFIDENCE_HIGH	Threshold for "high"	0.85
OGREP_CONFIDENCE_MEDIUM	Threshold for "medium"	0.70
OGREP_CONFIDENCE_LOW	Threshold for "low"	0.50

Smart Model Default:

• If OGREP_BASE_URL is set → defaults to nomic (local)
• Otherwise → defaults to text-embedding-3-small (OpenAI)

---

Multi-Repo Scope Management

Prevent cross-repo pollution:

Flag	Description
--db PATH	Custom database path
--profile NAME	Named profile (.ogrep/<name>/index.sqlite)
--global-cache	Use ~/.cache/ogrep/<hash>/index.sqlite
--repo-root PATH	Explicit repo root

---

Example Queries

# Find implementations
ogrep query "where is user authentication handled?" -n 10

# Find error handling
ogrep query "how are API errors handled?" -n 15 --rerank

# Find database operations
ogrep query "database connection and queries" -n 10

# Find specific patterns
ogrep query "recursive file scanning" -n 5

---

Documentation

• LOCAL_EMBEDDINGS_GUIDE.md — Local model setup, tuning, and troubleshooting
• QUICKSTART.md — Quick start guide
• CLAUDE.md — Developer guide for Claude Code
• WORD_ABOUT_SKILLUSE.md — Adapting CLAUDE.md for skill usage

---

Development

git clone https://github.com/gplv2/ogrep-marketplace.git
cd ogrep-marketplace
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,ast,rerank]"

make test    # Run tests (377 tests)
make lint    # Run linters
make check   # All checks

---

License

MIT