Skip to content

QUICK_REFERENCE.md

Latest commit

 

History

History
271 lines (215 loc) · 5.92 KB

QUICK_REFERENCE.md

File metadata and controls

271 lines (215 loc) · 5.92 KB

Quick Reference Card - Gemini Deep Agent

Setup (2 minutes)

pip install -r requirements.txt
cp .env.example .env
# Edit .env with your API keys

Minimal Example

import os
from langchain_google_genai import ChatGoogleGenerativeAI
from deepagents import create_deep_agent
from tavily import TavilyClient

# Setup
model = ChatGoogleGenerativeAI(model="gemini-2.5-flash")
tavily = TavilyClient(api_key=os.environ["TAVILY_API_KEY"])

# Tool
def search(query: str):
    return tavily.search(query)

# Agent
agent = create_deep_agent(
    tools=[search],
    system_prompt="You are a research expert.",
    model=model,
)

# Run
result = agent.invoke({
    "messages": [{"role": "user", "content": "Research AI agents"}]
})

print(result["messages"][-1].content)

Key Differences from Anthropic

# Anthropic (Original)
from langchain_anthropic import ChatAnthropic
model = ChatAnthropic(model="claude-3-5-sonnet-20241022")

# Google (This Project) ⭐
from langchain_google_genai import ChatGoogleGenerativeAI
model = ChatGoogleGenerativeAI(model="gemini-2.5-flash")

Configuration

model = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",           # Model ID
    api_key=os.environ.get("GOOGLE_API_KEY"),
    temperature=0.7,                    # 0=deterministic, 1=creative
    max_tokens=2048,                    # Output limit (optional)
    top_p=0.9,                          # Nucleus sampling (optional)
    top_k=40,                           # Top-k sampling (optional)
)

Built-in Agent Tools

Every deep agent automatically has:

write_todos(items)      # Plan tasks
read_file(path)         # Read files
write_file(path, content)
edit_file(path, old, new)
ls(dir)                 # List directory
grep(pattern, files)    # Search text
glob(pattern)           # Find files
execute(command)        # Run shell
task(prompt)            # Subagent

Custom Tools Pattern

def my_tool(param: str) -> str:
    """Tool description"""
    return "result"

agent = create_deep_agent(
    tools=[my_tool, search_tool],
    system_prompt="...",
    model=model,
)

Agent Invocation

result = agent.invoke({
    "messages": [{
        "role": "user",
        "content": "Your request here"
    }]
})

# Get response
response = result["messages"][-1].content
print(response)

Model Comparison

Factor Claude Gemini
Speed ⭐⭐⭐ ⭐⭐⭐⭐⭐
Cost ⭐⭐ ⭐⭐⭐⭐⭐
Reasoning ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐
Context ⭐⭐⭐⭐ ⭐⭐⭐⭐⭐

Typical Task Flow

User Request
    ↓
Agent Plans (write_todos)
    ↓
Search Information (search_tool)
    ↓
Organize Data (write_file)
    ↓
Analyze & Synthesize
    ↓
Generate Report
    ↓
Output

Environment Variables

# Required
export GOOGLE_API_KEY="AIzaSy..."
export TAVILY_API_KEY="tvly-..."

# Optional
export LANGSMITH_API_KEY="..."

File Examples

Simple Research

See: gemini_quickstart.py

Advanced Examples

See: advanced_examples.py

  • Comprehensive Researcher
  • News Analyst
  • Finance Researcher
  • Comparative Analyst
  • Tutorial Writer

Running Examples

# Basic research
python gemini_quickstart.py

# Advanced examples
python advanced_examples.py research
python advanced_examples.py news
python advanced_examples.py finance

Customization Checklist

  • Update system prompt for your domain
  • Add custom tools relevant to your task
  • Adjust temperature (0=deterministic, 1=creative)
  • Set max_tokens appropriate for output size
  • Test with sample inputs
  • Monitor costs and performance
  • Deploy when ready

Common Patterns

Research Task

prompt = """You are a researcher.
1. Plan subtopics with write_todos
2. Search each topic thoroughly
3. Save findings to files
4. Synthesize report"""

Analysis Task

prompt = """You are an analyst.
1. Search for relevant data
2. Organize in files
3. Perform analysis
4. Generate insights"""

Tutorial Creation

prompt = """You are a tutorial writer.
1. Plan structure with write_todos
2. Research best practices
3. Write detailed steps
4. Add examples and troubleshooting"""

Performance Tips

  1. Use Flash model - fastest & cheapest
  2. Limit search results - max_results=3-5
  3. Set appropriate temperature - 0.3-0.7 for consistency
  4. Use file tools - offload context
  5. Monitor with LangSmith - (optional) debug performance

Cost Estimation

For 1000 research tasks per month:

  • Claude: ~$45
  • Gemini 2.5 Flash: ~$1

Troubleshooting

Issue Solution
API key error Check export GOOGLE_API_KEY="..."
Module not found Run pip install -r requirements.txt
Rate limited Reduce queries or upgrade plan
Slow response Use Flash model, reduce results
Large context Use write_file to offload data

Useful Urls

Files in Project

deepagent2/
├── gemini_quickstart.py        # ⭐ Start here!
├── advanced_examples.py         # Advanced cases
├── GEMINI_QUICKSTART.md        # Full guide
├── COMPARISON.md               # Model comparison
├── README.md                   # Project overview
├── QUICK_REFERENCE.md          # This file
├── requirements.txt            # Dependencies
└── .env.example               # API key template

Next Steps

  1. Read: GEMINI_QUICKSTART.md
  2. Setup: Create .env file
  3. Run: python gemini_quickstart.py
  4. Explore: advanced_examples.py
  5. Customize: Modify for your use case
  6. Deploy: Use LangGraph for production

TL;DR: Install deps → Set API keys → Run gemini_quickstart.py → Customize