Search Tools Mastery: Combining rg, grepai, Serena, ast-grep, scip-search & lilmd

Master code search and documentation navigation by combining the right tools for maximum efficiency

Author: Florian BRUNIAUX | Contributions from Claude (Anthropic) Reading time: ~25 minutes Last updated: May 2026

---

Table of Contents

1. Quick Reference Matrix
2. Tool Comparison
3. Decision Tree
4. Combined Workflows
5. Real-World Scenarios
6. Performance Optimization
7. Common Pitfalls
8. Extended Toolkit: scip-search & lilmd

---

Quick Reference Matrix

I need to…	Use This Tool	Command Example
Find exact text	rg (Grep tool)	rg "authenticate" --type ts
Find by meaning	grepai	grepai search "user login flow"
Find function definition	Serena	serena find_symbol --name "login"
Find structural pattern	ast-grep	ast-grep "async function $F"
See who calls function	grepai	grepai trace callers "login"
Get file structure	Serena	serena get_symbols_overview
Refactor across files	Serena + ast-grep	Combined workflow
Explore unknown codebase	grepai → Serena	Discovery pattern
Find symbol refs without MCP (worktree-safe)	scip-search	scip-search refs AuthService.login
Navigate a large Markdown document	lilmd	lilmd read docs/arch.md "Authentication"

---

Tool Comparison

Complete Feature Matrix

Feature	rg (ripgrep)	grepai	Serena	ast-grep
Search Type	Regex/text	Semantic (meaning)	Symbol-aware	AST structure
Technology	Pattern matching	Embeddings (Ollama)	Symbol parsing	Abstract Syntax Tree
Speed	⚡ ~20ms	🐢 ~500ms	⚡ ~100ms	🕐 ~200ms
Setup	✅ None (built-in)	⚠️ Ollama + install	⚠️ MCP config	⚠️ npm install
Integration	✅ Native (Grep)	⚠️ MCP server	⚠️ MCP server	⚠️ Plugin
Privacy	✅ 100% local	✅ 100% local	✅ 100% local	✅ 100% local
Context needed	None	None	Project indexation	None
Languages	All (text)	All	TS/JS/Py/Rust/Go	TS/JS/Py/Rust/Go/C++
Call graph	❌ No	✅ Yes	❌ No	❌ No
Symbol tracking	❌ No	❌ No	✅ Yes	❌ No
Session memory	❌ No	❌ No	✅ Yes	❌ No
False positives	Medium	Low	Very low	Very low
Learning curve	Low	Medium	Low	High

Token Cost Comparison

Tool	Typical Query	Tokens Consumed	Results Returned
rg	”authenticate”	~500	Exact matches only
grepai	”auth flow”	~2000	Intent-based matches
Serena	find_symbol	~1000	Symbol + context
ast-grep	AST pattern	~1500	Structural matches

Key insight: rg is 4x more token-efficient but 10x less intelligent than semantic tools.

Extended Tool Reference

Two tools that address gaps in the core stack, covered in depth in § Extended Toolkit:

Tool	Category	vs existing stack
scip-search	Symbol index search (SCIP)	Like Serena but stateless, no MCP, worktree-safe
lilmd	Markdown section navigation	No equivalent in the stack (targets docs, not code)

---

Decision Tree

Level 1: What Do You Know?

Do you know the EXACT text/pattern?
│
├─ YES → Use rg (ripgrep)
│  ├─ Known function name: rg "createSession"
│  ├─ Known import: rg "import.*React"
│  └─ Known pattern: rg "async function"
│
└─ NO → Go to Level 2

Level 2: What Are You Looking For?

What's your search intent?
│
├─ "Find by MEANING/CONCEPT"
│  → Use grepai
│  └─ Example: grepai search "payment validation logic"
│
├─ "Find FUNCTION/CLASS definition"
│  → Use Serena
│  └─ Example: serena find_symbol --name "UserController"
│
├─ "Find by CODE STRUCTURE"
│  → Use ast-grep
│  └─ Example: async without error handling
│
└─ "Understand DEPENDENCIES"
   → Use grepai trace
   └─ Example: grepai trace callers "validatePayment"

Level 2 (Worktree / No MCP)

When running in a CI environment, a git worktree, or any context where MCP servers are unavailable:

Known symbol name, no MCP available?
│
└─ Use scip-search (pre-built SCIP index, millisecond cold start)
   └─ scip-search refs "AuthService.login" --format json

Level 3: Optimization

Found too many results?
│
├─ rg → Add --type filter or narrow path
├─ grepai → Add --path filter or use trace
├─ Serena → Filter by symbol type (function/class)
└─ ast-grep → Add constraints to pattern

---

Combined Workflows

Workflow 1: Exploring Unknown Codebase

Goal: Understand a new project quickly

Step-by-step:

# 1. SEMANTIC DISCOVERY (grepai)
# Find files related to authentication
grepai search "user authentication and session management"
# → Output: auth.service.ts, session.middleware.ts, user.controller.ts

# 2. STRUCTURAL OVERVIEW (Serena)
# Understand each file's structure
serena get_symbols_overview --file auth.service.ts
# → Output:
#   - class AuthService
#     - login(email, password)
#     - logout(sessionId)
#     - validateSession(token)

# 3. DEPENDENCY MAPPING (grepai trace)
# See how login is used
grepai trace callers "login"
# → Output: Called by UserController, ApiGateway, AdminPanel

# 4. EXACT SEARCH (rg)
# Find specific implementation details
rg "validateSession" --type ts -A 5
# → Output: Full function with 5 lines of context

Result: Complete understanding in 4 commands (vs 30+ file reads)

---

Workflow 2: Large-Scale Refactoring

Goal: Rename createSession → initializeUserSession across 50+ files

Step-by-step:

# 1. IMPACT ANALYSIS (grepai trace)
# Understand full scope
grepai trace callers "createSession"
# → Output: 47 callers across 23 files
grepai trace callees "createSession"
# → Output: Calls validateUser, createToken, storeSession

# 2. STRUCTURAL VALIDATION (ast-grep)
# Ensure consistent usage pattern
ast-grep "createSession($$$ARGS)"
# → Output: All invocations with their argument patterns

# 3. SYMBOL-AWARE REFACTORING (Serena)
# Precise renaming
serena find_symbol --name "createSession" --include-body true
# → Get exact definition + all references

serena replace_symbol_body \
  --name "createSession" \
  --new-name "initializeUserSession"
# → Rename across all files maintaining structure

# 4. VERIFICATION (rg)
# Confirm no old references remain
rg "createSession" --type ts
# → Should return 0 results

Result: Safe refactoring with full dependency awareness

---

Workflow 3: Security Audit

Goal: Find security vulnerabilities

Step-by-step:

# 1. SEMANTIC DISCOVERY (grepai)
# Find security-sensitive code
grepai search "SQL query construction"
grepai search "user input validation"
grepai search "password handling"

# 2. STRUCTURAL PATTERNS (ast-grep)
# Find specific vulnerability patterns

# SQL injection risks
ast-grep 'db.query(`${$VAR}`)'

# XSS risks
ast-grep 'innerHTML = $VAR'

# Missing error handling
ast-grep -p 'async function $F($$$) { $$$BODY }' \
  --without 'try { $$$TRY } catch'

# 3. DEPENDENCY TRACING (grepai)
# See where vulnerable code is called
grepai trace callers "executeQuery"
# → Identify all entry points

# 4. EXACT VERIFICATION (rg)
# Confirm findings
rg "innerHTML\s*=" --type ts
rg "password" --type ts | rg -v "hashed"

Result: Comprehensive security audit in minutes

---

Real-World Benchmarks

grepai vs grep (Janvier 2026)

Contexte: Benchmark sur Excalidraw (155k lignes TypeScript) Auteur: YoanDev (mainteneur de grepai - biais potentiel) Méthodologie: 5 questions de découverte de code identiques

Métrique	grep	grepai	Différence
Tool calls	139	62	-55%
Input tokens	51k	1.3k	-97%

À retenir: Recherche sémantique réduit drastiquement les tokens en identifiant les fichiers pertinents dès la première tentative, évitant l’exploration itérative.

Limitations:

• Benchmark par le mainteneur de l’outil
• Single-project validation (TypeScript only)
• Pas de validation indépendante à ce jour

Source: yoandev.co/grepai-benchmark

Note: Ce benchmark reflète l’état de janvier 2026. Les performances peuvent évoluer avec les mises à jour de Claude Code et grepai.

---

Workflow 4: Framework Migration

Goal: Migrate React class components → hooks

Step-by-step:

# 1. INVENTORY (ast-grep)
# Find all class components
ast-grep 'class $C extends React.Component'
# → Output: 34 components to migrate

# 2. DEPENDENCY ANALYSIS (grepai)
# Understand component relationships
for component in $(ast-grep 'class $C extends' --json | jq -r '.[].name'); do
  grepai trace callers "$component"
done
# → Build migration order (leaf components first)

# 3. PATTERN DETECTION (ast-grep)
# Identify lifecycle methods used
ast-grep 'componentDidMount() { $$$BODY }'
ast-grep 'componentWillReceiveProps($$$) { $$$BODY }'
# → Map to equivalent hooks

# 4. INCREMENTAL MIGRATION (Serena + ast-grep)
# Migrate one component at a time
serena find_symbol --name "UserProfile" --include-body true
# → Get full component code

# Use ast-grep to transform
ast-grep --rewrite \
  --from 'class $C extends React.Component' \
  --to 'const $C = () => { }'

# 5. VERIFICATION (rg + grepai)
# Ensure migration successful
rg "React.Component" --type tsx  # Should decrease
grepai search "component lifecycle methods"  # Find any missed

Result: Systematic migration with minimal breakage

---

Workflow 5: Performance Optimization

Goal: Identify and fix performance bottlenecks

Step-by-step:

# 1. HOTSPOT DISCOVERY (grepai)
# Find performance-critical code
grepai search "heavy computation or loops"
grepai search "database queries in loops"

# 2. PATTERN DETECTION (ast-grep)
# Find N+1 query patterns
ast-grep 'for ($$$) { await db.query($$$) }'

# Find missing memoization
ast-grep 'useMemo' --invert-match \
  --in 'const $VAR = $$$'

# 3. CALL GRAPH ANALYSIS (grepai trace)
# Find hot paths
grepai trace graph "renderUserList" --depth 3
# → Visualize dependency tree

# 4. SYMBOL TRACKING (Serena)
# Track function changes
serena write_memory "perf_baseline" \
  "renderUserList: 450ms avg"

# After optimization
serena write_memory "perf_optimized" \
  "renderUserList: 45ms avg (10x improvement)"

# 5. VERIFICATION (rg)
# Confirm optimizations applied
rg "useMemo|useCallback" --type tsx

Result: Data-driven performance improvements

---

Real-World Scenarios

Scenario 1: “I Don’t Know What I’m Looking For”

Problem: New project, no documentation, need to add feature

Solution: Semantic-first discovery

# Start broad with meaning
grepai search "user profile management"
# → Discover relevant files

# Then narrow with structure
serena get_symbols_overview --file user-profile.service.ts
# → Understand available functions

# Finally, exact search for details
rg "updateProfile" --type ts -C 3

---

Scenario 2: “This Function is Called from Everywhere”

Problem: Need to modify a function but worried about breaking things

Solution: Dependency mapping first

# 1. See all callers
grepai trace callers "calculateTotal"
# → 47 callers found

# 2. Analyze caller contexts
for file in $(grepai trace callers "calculateTotal" --json | jq -r '.[].file'); do
  serena get_symbols_overview --file "$file"
done

# 3. Identify safe vs risky call sites
ast-grep 'calculateTotal($ARGS)' --json
# → Group by argument patterns

# 4. Make change with confidence
# Now you know all impact points

---

Scenario 3: “Find All Code Doing X”

Problem: Need to apply consistent pattern across codebase

Solution: Combine semantic + structural

# Example: Find all error handling code

# 1. Semantic discovery
grepai search "error handling and exception management"

# 2. Structural patterns
ast-grep 'try { $$$TRY } catch ($ERR) { $$$CATCH }'
ast-grep 'throw new Error($MSG)'

# 3. Verify consistency
rg "catch\s*\(" --type ts | wc -l
# Compare with ast-grep count to find anomalies

---

Scenario 4: “I Need to Understand This Module”

Problem: Complex module with unclear responsibilities

Solution: Multi-tool analysis

# 1. Get symbol overview (Serena)
serena get_symbols_overview --file payment.module.ts
# → See all exports, classes, functions

# 2. Understand dependencies (grepai)
grepai trace callees "PaymentModule"
# → What does this module use?

grepai trace callers "PaymentModule"
# → Who uses this module?

# 3. Find implementation patterns (ast-grep)
ast-grep 'export class $C' --file payment.module.ts
ast-grep 'async $METHOD($$$)' --file payment.module.ts

# 4. Read specific implementations (rg)
rg "processPayment" --type ts -A 20

---

Performance Optimization

Choosing the Fastest Tool

General Rules:

1. Known exact text → Always use rg first
2. Unknown exact text → Use grepai, then rg for verification
3. Refactoring → Serena for symbol safety
4. Large migrations → ast-grep for structural precision

Performance Benchmarks

Test: Find authentication code in 500k line codebase

Strategy	Time	Results Quality
rg “auth” only	0.2s	5000+ false positives
grepai “auth” only	2.5s	50 relevant results
grepai → rg (combined)	2.7s	50 relevant, verified
Serena symbols only	1.5s	12 auth functions
ast-grep patterns	3.0s	8 auth flows

Winner: Serena symbols (fastest + high quality) for known function names

Parallelization Strategy

For large codebases (>100k lines):

# Run searches in parallel

# Terminal 1: Semantic discovery
grepai search "authentication flow" > /tmp/grepai-results.json &

# Terminal 2: Symbol indexing
serena get_symbols_overview --file src/**/*.ts > /tmp/symbols.json &

# Terminal 3: Pattern detection
ast-grep 'async function $F' --json > /tmp/ast-results.json &

# Wait for all, then combine results
wait
jq -s '.[0] + .[1] + .[2]' \
  /tmp/grepai-results.json \
  /tmp/symbols.json \
  /tmp/ast-results.json

---

Common Pitfalls

Pitfall 1: Using Semantic Search for Exact Matches

❌ Wrong:

grepai search "createSession"  # Slow, overkill

✅ Right:

rg "createSession" --type ts  # Fast, precise

Rule: If you know the exact text, never use semantic search.

---

Pitfall 2: Using rg for Conceptual Search

❌ Wrong:

rg "auth.*login.*session" --type ts  # Misses variations

✅ Right:

grepai search "authentication and session management"

Rule: Regex doesn’t understand meaning, use semantic tools.

---

Pitfall 3: Ignoring Call Graph Before Refactoring

❌ Wrong:

# Directly refactor without checking callers
rg "oldFunction" --type ts | sed 's/oldFunction/newFunction/g'

✅ Right:

# Check impact first
grepai trace callers "oldFunction"
# See 47 callers across 23 files
# Then plan refactoring strategy

Rule: Always trace dependencies before modifying shared code.

---

Pitfall 4: Not Combining Tools

❌ Wrong:

# Use only one tool for complex task
ast-grep 'async function $F' --json | jq '.[].file' | xargs -I {} vim {}
# Blindly edit without understanding context

✅ Right:

# Combine for full understanding
ast-grep 'async function $F' --json > /tmp/async.json
for file in $(jq -r '.[].file' /tmp/async.json); do
  serena get_symbols_overview --file "$file"  # Context
  grepai trace callers "$(jq -r '.[].name' /tmp/async.json)"  # Usage
done

Rule: Complex tasks need multiple perspectives.

---

Pitfall 5: Over-Engineering Simple Searches

❌ Wrong:

# Setup grepai + Ollama just to find a TODO comment
grepai search "TODO comments in the code"

✅ Right:

rg "TODO" --type ts

Rule: Use the simplest tool that works.

---

Tool Selection Cheatsheet

Quick Decision Matrix

Your Situation	Use This	Not This
”Find function login”	rg “login”	grepai search “login"
"Find login-related code”	grepai “login flow”	rg “login.*"
"Rename function safely”	Serena find_symbol	rg + sed
”Who calls this function?“	grepai trace callers	rg + grep
”Get file structure”	Serena overview	rg “class|function"
"Find async without try/catch”	ast-grep	rg “async.*{"
"Migrate React classes”	ast-grep	rg + manual
”Find TODOs”	rg “TODO”	Any other tool

---

Extended Toolkit: scip-search & lilmd

Two CLI tools that address gaps in the rg/grepai/Serena stack.

scip-search

scip-search queries pre-built SCIP (Sourcegraph Code Intelligence Protocol) symbol indexes. Where grepai searches by semantic meaning and Serena requires a live MCP connection, scip-search operates against a static binary index with millisecond cold starts.

Attribute	Details
Source	github.com/liza-mas/scip-search
Install	curl -fsSL https://raw.githubusercontent.com/liza-mas/scip-search/main/install.sh | bash
Index format	SCIP (Go, TypeScript, Python, Java, Rust, and others)
Output	One-line text, JSON, or location-only

Workflow: scip-search replaces the 5-10 rg/read round-trips typical for symbol discovery.

# Step 1: generate SCIP index once per language
scip-typescript index --output index.scip

# Step 2: find a symbol definition
scip-search find AuthService

# Step 3: get all references with line numbers
scip-search refs AuthService.login --format json

# Step 4: read only the returned line ranges

vs Serena: Serena connects to a language server and has session memory. scip-search is stateless and works against a snapshot: no MCP, no persistent process. This makes it reliable in worktrees and ephemeral CI environments where Serena’s LSP backend may not be available.

vs grepai: grepai finds by semantic intent (“payment validation logic”). scip-search finds by exact or near-exact symbol identifier. They work in sequence: grepai discovers the concept, scip-search confirms the symbol.

Worktree compatibility: indexes are per-repository files with no shared state. Running scip-typescript index inside a worktree produces a local index for that worktree.

---

lilmd

lilmd treats Markdown files as databases. It returns a table of contents with line ranges and enables targeted section reads. An agent reading a 2,000-line guide fetches one section in a single call instead of reading the full file.

Attribute	Details
Source	github.com/molefrog/lilmd
Install	npm install -g lilmd
Runtime	Node or Bun

Key commands:

# TOC with line ranges (inclusive, 1-indexed)
lilmd docs/architecture.md

# Read a section by name (fuzzy match by default)
lilmd read docs/architecture.md "Authentication"

# Read a nested section
lilmd read docs/architecture.md "Security > JWT"

# Exact match (prefix with =)
lilmd read docs/architecture.md "=Authentication Flow"

Agent-specific value: the TOC output contains line ranges for each heading. An agent parses these and requests only the relevant section rather than loading the full file. The natural pipeline is: rg (find which file) then lilmd (get TOC with ranges) then Read lines:N-M (load the specific section). This also works for CHANGELOG.md, large README files, and knowledge base documents.

Worktree compatibility: stateless, no index, runs per file. Works anywhere.

---

Setup Priority

Recommended Setup Order:

1. Start: rg (already built-in with Grep tool) ✅
2. Next: Serena MCP (symbol awareness, session memory)
3. Then: grepai (semantic search + call graph)
4. If worktrees are part of your workflow: scip-search (stateless symbol lookup, no MCP required)
5. For large documentation: lilmd (targeted Markdown section reads, no setup)
6. Finally: ast-grep (structural patterns, large refactoring)

Rationale: 90% of searches work with rg + Serena. Add grepai for semantic needs. Add scip-search for worktree or CI environments where MCP is unavailable. Add ast-grep only for large-scale refactoring.

---

Summary: The 6-Tool Toolkit

┌─────────────────────────────────────────────────────────┐
│                   SEARCH TOOL MASTERY                   │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  rg (ripgrep)     →  Fast, exact text matching         │
│  ├─ Use: 90% of searches                               │
│  └─ Speed: ~20ms                                        │
│                                                         │
│  grepai           →  Semantic + Call graph             │
│  ├─ Use: Concept discovery, dependency tracing         │
│  └─ Speed: ~500ms (finds what rg cannot)               │
│                                                         │
│  Serena           →  Symbol-aware + Session memory     │
│  ├─ Use: Refactoring, structure understanding          │
│  └─ Speed: ~100ms                                       │
│                                                         │
│  ast-grep         →  AST structural patterns           │
│  ├─ Use: Large migrations, complex patterns            │
│  └─ Speed: ~200ms                                       │
│                                                         │
│  scip-search      →  Symbol index (stateless, SCIP)   │
│  ├─ Use: CI / worktrees / no MCP environments          │
│  └─ Speed: ~5ms cold start                             │
│                                                         │
│  lilmd            →  Markdown section navigation       │
│  ├─ Use: Large docs, TOC + line ranges per section     │
│  └─ Speed: instant, no index                           │
│                                                         │
│  ═══════════════════════════════════════════════════   │
│                                                         │
│  Master the combination, not individual tools.          │
│  Each tool has a sweet spot. Use the right one.        │
│                                                         │
└─────────────────────────────────────────────────────────┘

---

Further Reading

• Serena MCP Guide
• grepai Documentation
• ast-grep Patterns Skill
• Architecture: Grep vs RAG History
• scip-search GitHub
• lilmd GitHub

---

Last updated: May 2026 Compatible with: Claude Code 2.1.7+