josondev's picture
Update veryfinal.py
d52b24c verified
raw
history blame
12.8 kB
"""LangGraph Agent with Best Free Models and Minimal Rate Limits"""
import os, time, random
from dotenv import load_dotenv
from typing import List, Dict, Any, TypedDict, Annotated
import operator
# LangGraph imports
from langgraph.graph import START, StateGraph, MessagesState
from langgraph.prebuilt import tools_condition
from langgraph.prebuilt import ToolNode
from langgraph.checkpoint.memory import MemorySaver
# LangChain imports
from langchain_core.messages import SystemMessage, HumanMessage
from langchain_core.tools import tool
from langchain_groq import ChatGroq
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_nvidia_ai_endpoints import ChatNVIDIA
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_community.document_loaders import WikipediaLoader, ArxivLoader
from langchain_community.vectorstores import FAISS
from langchain_nvidia_ai_endpoints import NVIDIAEmbeddings
from langchain.tools.retriever import create_retriever_tool
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import JSONLoader
load_dotenv()
# Advanced Rate Limiter with Exponential Backoff
class AdvancedRateLimiter:
def __init__(self, requests_per_minute: int, provider_name: str):
self.requests_per_minute = requests_per_minute
self.provider_name = provider_name
self.request_times = []
self.consecutive_failures = 0
def wait_if_needed(self):
current_time = time.time()
# Clean old requests (older than 1 minute)
self.request_times = [t for t in self.request_times if current_time - t < 60]
# Check if we need to wait
if len(self.request_times) >= self.requests_per_minute:
wait_time = 60 - (current_time - self.request_times[0]) + random.uniform(2, 8)
time.sleep(wait_time)
# Add exponential backoff for consecutive failures
if self.consecutive_failures > 0:
backoff_time = min(2 ** self.consecutive_failures, 60) + random.uniform(1, 3)
time.sleep(backoff_time)
# Record this request
self.request_times.append(current_time)
def record_success(self):
self.consecutive_failures = 0
def record_failure(self):
self.consecutive_failures += 1
# Initialize rate limiters based on search results
# Gemini 2.0 Flash-Lite: 30 RPM (highest free tier)
gemini_limiter = AdvancedRateLimiter(requests_per_minute=25, provider_name="Gemini") # Conservative
# Groq: Typically 30 RPM for free tier
groq_limiter = AdvancedRateLimiter(requests_per_minute=25, provider_name="Groq") # Conservative
# NVIDIA: Typically 5 RPM for free tier
nvidia_limiter = AdvancedRateLimiter(requests_per_minute=4, provider_name="NVIDIA") # Very conservative
# Initialize LLMs with best models and minimal rate limits
def get_best_models():
"""Get the best models with lowest rate limits"""
# Gemini 2.0 Flash-Lite - Best rate limit (30 RPM) with good performance
gemini_llm = ChatGoogleGenerativeAI(
model="gemini-2.0-flash-lite", # Best rate limit from search results
api_key=os.getenv("GOOGLE_API_KEY"),
temperature=0,
max_output_tokens=4000
)
# Groq Llama 3.3 70B - Fast and capable
groq_llm = ChatGroq(
model="llama-3.3-70b-versatile",
api_key=os.getenv("GROQ_API_KEY"),
temperature=0,
max_tokens=4000
)
# NVIDIA Llama 3.1 70B - Good for specialized tasks
nvidia_llm = ChatNVIDIA(
model="meta/llama-3.1-70b-instruct",
api_key=os.getenv("NVIDIA_API_KEY"),
temperature=0,
max_tokens=4000
)
return {
"gemini": gemini_llm,
"groq": groq_llm,
"nvidia": nvidia_llm
}
# Fallback strategy with rate limit handling
class ModelFallbackManager:
def __init__(self):
self.models = get_best_models()
self.limiters = {
"gemini": gemini_limiter,
"groq": groq_limiter,
"nvidia": nvidia_limiter
}
self.fallback_order = ["gemini", "groq", "nvidia"] # Order by rate limit capacity
def invoke_with_fallback(self, messages, max_retries=3):
"""Try models in order with rate limiting and fallbacks"""
for provider in self.fallback_order:
limiter = self.limiters[provider]
model = self.models[provider]
for attempt in range(max_retries):
try:
# Apply rate limiting
limiter.wait_if_needed()
# Try to invoke the model
response = model.invoke(messages)
limiter.record_success()
return response
except Exception as e:
error_msg = str(e).lower()
# Check if it's a rate limit error
if any(keyword in error_msg for keyword in ['rate limit', '429', 'quota', 'too many requests']):
limiter.record_failure()
wait_time = (2 ** attempt) + random.uniform(10, 30)
time.sleep(wait_time)
continue
else:
# Non-rate limit error, try next provider
break
# If all providers fail
raise Exception("All model providers failed or hit rate limits")
# Custom Tools
@tool
def multiply(a: int, b: int) -> int:
"""Multiply two numbers."""
return a * b
@tool
def add(a: int, b: int) -> int:
"""Add two numbers."""
return a + b
@tool
def subtract(a: int, b: int) -> int:
"""Subtract two numbers."""
return a - b
@tool
def divide(a: int, b: int) -> float:
"""Divide two numbers."""
if b == 0:
raise ValueError("Cannot divide by zero.")
return a / b
@tool
def modulus(a: int, b: int) -> int:
"""Get the modulus of two numbers."""
return a % b
@tool
def wiki_search(query: str) -> str:
"""Search Wikipedia for a query and return maximum 2 results."""
try:
time.sleep(random.uniform(1, 3))
search_docs = WikipediaLoader(query=query, load_max_docs=2).load()
formatted_search_docs = "\n\n---\n\n".join(
[
f'<Document source="{doc.metadata["source"]}" page="{doc.metadata.get("page", "")}"/>\n{doc.page_content}\n</Document>'
for doc in search_docs
])
return formatted_search_docs
except Exception as e:
return f"Wikipedia search failed: {str(e)}"
@tool
def web_search(query: str) -> str:
"""Search Tavily for a query and return maximum 3 results."""
try:
time.sleep(random.uniform(2, 5))
search_docs = TavilySearchResults(max_results=3).invoke(query=query)
formatted_search_docs = "\n\n---\n\n".join(
[
f'<Document source="{doc.get("url", "")}" />\n{doc.get("content", "")}\n</Document>'
for doc in search_docs
])
return formatted_search_docs
except Exception as e:
return f"Web search failed: {str(e)}"
@tool
def arvix_search(query: str) -> str:
"""Search Arxiv for a query and return maximum 3 result."""
try:
time.sleep(random.uniform(1, 4))
search_docs = ArxivLoader(query=query, load_max_docs=3).load()
formatted_search_docs = "\n\n---\n\n".join(
[
f'<Document source="{doc.metadata["source"]}" page="{doc.metadata.get("page", "")}"/>\n{doc.page_content[:1000]}\n</Document>'
for doc in search_docs
])
return formatted_search_docs
except Exception as e:
return f"ArXiv search failed: {str(e)}"
# Setup FAISS vector store
def setup_faiss_vector_store():
"""Setup FAISS vector database from JSONL metadata"""
try:
jq_schema = """
{
page_content: .Question,
metadata: {
task_id: .task_id,
Level: .Level,
Final_answer: ."Final answer",
file_name: .file_name,
Steps: .["Annotator Metadata"].Steps,
Number_of_steps: .["Annotator Metadata"]["Number of steps"],
How_long: .["Annotator Metadata"]["How long did this take?"],
Tools: .["Annotator Metadata"].Tools,
Number_of_tools: .["Annotator Metadata"]["Number of tools"]
}
}
"""
json_loader = JSONLoader(file_path="metadata.jsonl", jq_schema=jq_schema, json_lines=True, text_content=False)
json_docs = json_loader.load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=512, chunk_overlap=200)
json_chunks = text_splitter.split_documents(json_docs)
embeddings = NVIDIAEmbeddings(
model="nvidia/nv-embedqa-e5-v5",
api_key=os.getenv("NVIDIA_API_KEY")
)
vector_store = FAISS.from_documents(json_chunks, embeddings)
return vector_store
except Exception as e:
print(f"FAISS vector store setup failed: {e}")
return None
# Load system prompt
try:
with open("system_prompt.txt", "r", encoding="utf-8") as f:
system_prompt = f.read()
except FileNotFoundError:
system_prompt = """You are a helpful assistant tasked with answering questions using a set of tools.
Now, I will ask you a question. Report your thoughts, and finish your answer with the following template:
FINAL ANSWER: [YOUR FINAL ANSWER].
YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list of numbers and/or strings."""
sys_msg = SystemMessage(content=system_prompt)
# Setup vector store and retriever
vector_store = setup_faiss_vector_store()
if vector_store:
retriever = vector_store.as_retriever(search_type="similarity", search_kwargs={"k": 3})
retriever_tool = create_retriever_tool(
retriever=retriever,
name="Question_Search",
description="A tool to retrieve similar questions from a vector store.",
)
else:
retriever_tool = None
# All tools
all_tools = [multiply, add, subtract, divide, modulus, wiki_search, web_search, arvix_search]
if retriever_tool:
all_tools.append(retriever_tool)
# Build graph function with fallback manager
def build_graph(provider="groq"):
"""Build the LangGraph with rate limiting and fallbacks"""
fallback_manager = ModelFallbackManager()
# Create a wrapper LLM that uses fallback manager
class FallbackLLM:
def bind_tools(self, tools):
self.tools = tools
return self
def invoke(self, messages):
return fallback_manager.invoke_with_fallback(messages)
llm_with_tools = FallbackLLM().bind_tools(all_tools)
# Node functions
def assistant(state: MessagesState):
"""Assistant node with fallback handling"""
return {"messages": [llm_with_tools.invoke(state["messages"])]}
def retriever_node(state: MessagesState):
"""Retriever node"""
if vector_store and len(state["messages"]) > 0:
try:
similar_questions = vector_store.similarity_search(state["messages"][-1].content, k=1)
if similar_questions:
example_msg = HumanMessage(
content=f"Here I provide a similar question and answer for reference: \n\n{similar_questions[0].page_content}",
)
return {"messages": [sys_msg] + state["messages"] + [example_msg]}
except Exception as e:
print(f"Retriever error: {e}")
return {"messages": [sys_msg] + state["messages"]}
# Build graph
builder = StateGraph(MessagesState)
builder.add_node("retriever", retriever_node)
builder.add_node("assistant", assistant)
builder.add_node("tools", ToolNode(all_tools))
builder.add_edge(START, "retriever")
builder.add_edge("retriever", "assistant")
builder.add_conditional_edges("assistant", tools_condition)
builder.add_edge("tools", "assistant")
# Compile graph with memory
memory = MemorySaver()
return builder.compile(checkpointer=memory)
# Test
if __name__ == "__main__":
question = "What are the names of the US presidents who were assassinated?"
graph = build_graph()
messages = [HumanMessage(content=question)]
config = {"configurable": {"thread_id": "test_thread"}}
result = graph.invoke({"messages": messages}, config)
for m in result["messages"]:
m.pretty_print()