osgrep: Semantic Code Search

Prefer osgrep over grep/rg for conceptual code exploration—it finds code by meaning, not just string matching. For exact identifier or literal string searches, grep/rg remains appropriate.

Overview

osgrep is a natural-language semantic code search tool that finds code by concept rather than keyword matching. Unlike grep which matches literal strings, osgrep understands code semantics using local AI embeddings.

Version 0.5.16 (Dec 2025) highlights:

• skeleton command: Compress files to function/class signatures (~85% token reduction)

• trace command: Show who calls/what calls for any symbol (call graph)

• symbols command: List all indexed symbols with definitions

• doctor command: Health/integrity verification

• list command: Display all indexed repositories

• Per-project .osgrep/ directories (no longer global ~/.osgrep/data )

• V2 architecture with improved performance (~20% token savings, ~30% speedup)

• Go language support

• --reset flag for clean re-indexing

• ColBERT reranking for better result relevance

• Role detection: distinguishes orchestration logic from type definitions

• Split searching: separate "Code" and "Docs" indices

When to use osgrep:

• Exploring unfamiliar codebases ("where is the auth logic?")

• Finding conceptual patterns ("show me error handling")

• Locating cross-cutting concerns ("all database migrations")

• User explicitly asks to search code semantically

When to use traditional tools:

• Searching for exact strings or identifiers (use Grep )

• Finding files by name pattern (use Glob )

• Already know the exact location (use Read )

Quick Start

Run osgrep from within the project directory, since it uses per-project .osgrep/ indexes.

cd /path/to/project # REQUIRED: cd into the project first osgrep "your query" # Now search works

Basic Search

osgrep "your semantic query" osgrep search "your query" path/to/scope # Scope to subdirectory osgrep skeleton src/file.py # Compress file to signatures osgrep trace functionName # Show call graph osgrep symbols # List all symbols

Examples:

osgrep "user registration flow" osgrep "webhook signature validation" osgrep "database transaction handling" osgrep "how are plugins loaded" packages/src

Output Format

Returns results in this format:

IMPLEMENTATION path/to/file:line Score: 0.95

Preamble: [code snippet or content preview] ...

• IMPLEMENTATION: Tag indicating the type of match

• Score: Relevance score (0-1, higher is better)

• ...: Truncation marker—snippet is incomplete, use Read for full context

Search Strategy

For Architectural/System-Level Questions

Use for: auth, integrations, file watching, cross-cutting concerns

Search broadly first to map the landscape:

osgrep "authentication authorization checks"

Survey the results - look for patterns across multiple files:

• Are checks in middleware? Decorators? Multiple services?

• Do file paths suggest different layers (gateway, handlers, utils)?

Read strategically - pick 2-4 files that represent different aspects:

• Read the main entry point

• Read representative middleware/util files

• Follow imports if architecture is unclear

Refine with specific searches if one aspect is unclear:

osgrep "session validation logic" osgrep "API authentication middleware"

For Targeted Implementation Details

Use for: specific function, algorithm, single feature

Search specifically about the precise logic:

osgrep "logic for merging user and default configuration"

Evaluate the semantic match:

• Does the snippet look relevant?

• If it ends in ... or cuts off mid-logic, read the file

One search, one read: Use osgrep to pinpoint the best file, then read it fully.

CLI Reference

Search Options

Control result count:

osgrep "validation logic" -m 20 # Max 20 results total (default: 10) osgrep "validation logic" --per-file 3 # Up to 3 matches per file (default: 1)

Output formats:

osgrep "API endpoints" --compact # File paths only osgrep "API endpoints" --content # Full chunk content (not just snippets) osgrep "API endpoints" --scores # Show relevance scores osgrep "API endpoints" --plain # Disable ANSI colors

Sync before search:

osgrep "validation logic" -s # Sync files to index before searching osgrep "validation logic" -d # Dry run (show what would sync)

Index Management

osgrep index # Incremental update osgrep index -r # Full re-index from scratch (--reset) osgrep index -p /path/to/repo # Index a specific directory osgrep index -d # Preview what would be indexed (--dry-run)

Advanced Commands (v0.5+)

Skeleton - Compress files to signatures:

osgrep skeleton src/server.py # Show function/class signatures only osgrep skeleton src/server.py --no-summary # Omit call/complexity summaries osgrep skeleton "auth logic" -l 5 # Query mode: skeleton of top 5 matching files

Output shows: function signatures with # → calls | C:N | ORCH summaries inside bodies.

Trace - Show call graph:

osgrep trace handleRequest # Who calls this? What does it call?

Symbols - List all indexed symbols:

osgrep symbols # All symbols (default limit: 20) osgrep symbols "Request" # Filter by pattern osgrep symbols -p src/api/ -l 50 # Filter by path, increase limit

Other Commands

osgrep list # Show all indexed repositories osgrep doctor # Check health and configuration osgrep setup # Pre-download models (~150MB) osgrep serve # Run background daemon (port 4444) osgrep serve -p 8080 # Custom port (or OSGREP_PORT=8080) osgrep serve -b # Run in background (--background) osgrep serve status # Check if daemon is running osgrep serve stop # Stop daemon osgrep serve stop --all # Stop all daemons

Serve endpoints:

• GET /health

• Health check

• POST /search

• Search with { query, limit, path, rerank }

• Lock file: .osgrep/server.json with port /pid

Claude Code Integration

osgrep install-claude-code # Install as Claude Code plugin osgrep install-opencode # Install for Opencode

Both plugins automatically manage the background server lifecycle during sessions.

Common Search Patterns

Architecture Exploration

Mental processes (Open Souls / Daimonic)

osgrep "mental processes that orchestrate conversation flow" osgrep "subprocesses that learn about the user" osgrep "cognitive steps using structured output"

React/Next.js

osgrep "where do we fetch data in components?" osgrep "custom hooks for API calls" osgrep "protected route implementation"

Backend

osgrep "request validation middleware" osgrep "authentication flow" osgrep "rate limiting logic"

Business Logic

osgrep "payment processing" osgrep "notification sending" osgrep "user permission checks" osgrep "order fulfillment workflow"

Cross-Cutting Concerns

osgrep "error handling patterns" osgrep "logging configuration" osgrep "database migrations" osgrep "environment variable usage"

Tips for Effective Queries

Trust the Semantics

You don't need exact names. Conceptual queries work better:

Good - conceptual

osgrep "how does the server start" osgrep "component state management"

Less effective - too literal

osgrep "server.init" osgrep "useState"

Be Specific

Too vague

osgrep "code"

Clear intent

osgrep "user registration validation logic"

Use Natural Language

osgrep "how do we handle payment failures?" osgrep "what happens when a webhook arrives?" osgrep "where is user input sanitized?"

Watch for Distributed Patterns

If results span 5+ files in different directories, the feature is likely architectural—survey before diving deep.

Don't Over-Rely on Snippets

For architectural questions, snippets are signposts, not answers. Read the key files.

Technical Details

• 100% Local: Uses transformers.js embeddings (no remote API calls)

• Auto-Isolated: Each repo gets its own index in .osgrep/ directory (v0.5+)

• Adaptive Performance: Bounded concurrency keeps system responsive

• Index Location: .osgrep/ in project root (was ~/.osgrep/data/ in v0.4.x)

• Model Download: ~150MB on first run (osgrep setup to pre-download)

• Chunking Strategy: Tree-sitter parses code into function/class boundaries

• Deduplication: Identical code blocks are deduplicated

• Dual Channels: Separate "Code" and "Docs" indices with ColBERT reranking

• Structural Boosting: Functions/classes prioritized over test files

• Skeleton Compression: ~85% token reduction when viewing file structure

Troubleshooting

"Still Indexing..." message:

• Index is ongoing. Results will be partial until complete.

• Alert the user and ask if they wish to proceed.

Slow first search:

• Expected—indexing takes 30-60s for medium repos

• Use osgrep setup to pre-download models

Index out of date:

• Run osgrep index to refresh

• Run osgrep index --reset for a complete re-index

• osgrep usually auto-detects changes

Installation issues:

osgrep doctor # Diagnose problems npm install -g osgrep # Reinstall if needed

No results found:

• Try broader queries ("authentication" vs "JWT middleware")

• Ensure index is up to date (osgrep index )

• Verify you're in the correct repository directory