search_agent / search_agent.py
CyranoB's picture
Original code
9c3709d
raw
history blame
12.8 kB
"""search_agent.py
Usage:
search_agent.py
[--domain=domain]
[--provider=provider]
[--temperature=temp]
[--max_pages=num]
SEARCH_QUERY
search_agent.py --version
Options:
-h --help Show this screen.
--version Show version.
-d domain --domain=domain Limit search to a specific domain
-t temp --temperature=temp Set the temperature of the LLM [default: 0.0]
-p provider --provider=provider Use a specific LLM (choices: bedrock,openai,groq) [default: openai]
-m num --max_pages=num Max number of pages to retrieve [default: 10]
"""
import json
import os
from concurrent.futures import ThreadPoolExecutor
from urllib.parse import quote
from bs4 import BeautifulSoup
from docopt import docopt
import dotenv
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.schema import SystemMessage, HumanMessage
from langchain.callbacks import LangChainTracer
from langchain_groq import ChatGroq
from langchain_openai import ChatOpenAI
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores.faiss import FAISS
from langchain_community.chat_models.bedrock import BedrockChat
from langsmith import Client
import requests
from rich.console import Console
from rich.rule import Rule
from rich.markdown import Markdown
def get_chat_llm(provider, temperature=0.0):
console.log(f"Using provider {provider} with temperature {temperature}")
match provider:
case 'bedrock':
chat_llm = BedrockChat(
credentials_profile_name=os.getenv('CREDENTIALS_PROFILE_NAME'),
model_id="anthropic.claude-3-sonnet-20240229-v1:0",
model_kwargs={"temperature": temperature },
)
case 'openai':
chat_llm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=temperature)
case 'groq':
chat_llm = ChatGroq(model_name = 'mixtral-8x7b-32768', temperature=temperature)
case _:
raise ValueError(f"Unknown LLM provider {provider}")
return chat_llm
def optimize_search_query(query):
messages = [
SystemMessage(
content="""
You are a serach query optimizer specialist.
Rewrite the user's question using only the most important keywords. Remove extra words.
Tips:
Identify the key concepts in the question
Remove filler words like "how to", "what is", "I want to"
Removed style such as "in the style of", "engaging", "short", "long"
Remove lenght instruction (example: essay, article, letter, blog, post, blogpost, etc)
Keep it short, around 3-7 words total
Put the most important keywords first
Remove formatting instructions
Remove style instructions (exmaple: in the style of, engaging, short, long)
Remove lenght instruction (example: essay, article, letter, etc)
Example:
Question: How do I bake chocolate chip cookies from scratch?
Search query: chocolate chip cookies recipe from scratch
Example:
Question: I would like you to show me a time line of Marie Curie life. Show results as a markdown table
Search query: Marie Curie timeline
Example:
Question: I would like you to write a long article on nato vs russia. Use know geopolical frameworks.
Search query: geopolitics nato russia
Example:
Question: Write a engaging linkedin post about Andrew Ng
Search query: Andrew Ng
Example:
Question: Write a short artible about the solar system in the style of Carl Sagan
Search query: solar system
Example:
Question: Should I use Kubernetes? Answer in the style of Gilfoyde from the TV show Silicon Valley
Search query: Kubernetes decision
Example:
Question: biography of napoleon. include a table with the major events.
Search query: napoleon biography events
"""
),
HumanMessage(
content=f"""
Questions: {query}
Search query:
"""
),
]
response = chat.invoke(messages, config={"callbacks": callbacks})
return response.content
def get_sources(query, max_pages=10, domain=None):
search_query = query
if domain:
search_query += f" site:{domain}"
url = f"https://api.search.brave.com/res/v1/web/search?q={quote(search_query)}&count={max_pages}"
headers = {
'Accept': 'application/json',
'Accept-Encoding': 'gzip',
'X-Subscription-Token': os.getenv("BRAVE_SEARCH_API_KEY")
}
try:
response = requests.get(url, headers=headers)
if response.status_code != 200:
raise Exception(f"HTTP error! status: {response.status_code}")
json_response = response.json()
if 'web' not in json_response or 'results' not in json_response['web']:
raise Exception('Invalid API response format')
final_results = [{
'title': result['title'],
'link': result['url'],
'snippet': result['description'],
'favicon': result.get('profile', {}).get('img', '')
} for result in json_response['web']['results']]
return final_results
except Exception as error:
#console.log('Error fetching search results:', error)
raise
def fetch_with_timeout(url, timeout=8):
try:
response = requests.get(url, timeout=timeout)
response.raise_for_status()
return response
except requests.RequestException as error:
#console.log(f"Skipping {url}! Error: {error}")
return None
def extract_main_content(html):
try:
soup = BeautifulSoup(html, 'html.parser')
for element in soup(["script", "style", "head", "nav", "footer", "iframe", "img"]):
element.extract()
main_content = ' '.join(soup.body.get_text().split())
return main_content
except Exception as error:
#console.log(f"Error extracting main content: {error}")
return None
def process_source(source):
response = fetch_with_timeout(source['link'], 8)
if response:
html = response.text
main_content = extract_main_content(html)
return {**source, 'html': main_content}
return None
def get_links_contents(sources):
with ThreadPoolExecutor() as executor:
results = list(executor.map(process_source, sources))
# Filter out None results
return [result for result in results if result is not None]
def process_and_vectorize_content(
contents,
query,
text_chunk_size=1000,
text_chunk_overlap=200,
number_of_similarity_results=5
):
"""
Process and vectorize content using Langchain.
Args:
contents (list): List of dictionaries containing 'title', 'link', and 'html' keys.
query (str): Query string for similarity search.
text_chunk_size (int): Size of each text chunk.
text_chunk_overlap (int): Overlap between text chunks.
number_of_similarity_results (int): Number of most similar results to return.
Returns:
list: List of most similar documents.
"""
documents = []
for content in contents:
if content['html']:
try:
# Split text into chunks
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=text_chunk_size,
chunk_overlap=text_chunk_overlap
)
texts = text_splitter.split_text(content['html'])
# Create metadata for each text chunk
metadatas = [{'title': content['title'], 'link': content['link']} for _ in range(len(texts))]
# Create vector store
embeddings = OpenAIEmbeddings()
docsearch = FAISS.from_texts(texts, embedding=embeddings, metadatas=metadatas)
# Perform similarity search
docs = docsearch.similarity_search(query, k=number_of_similarity_results)
doc_dicts = [{'page_content': doc.page_content, 'metadata': doc.metadata} for doc in docs]
documents.extend(doc_dicts)
except Exception as e:
console.log(f"[gray]Error processing content for {content['link']}: {e}")
return documents
def answer_query_with_sources(query, relevant_docs):
messages = [
SystemMessage(
content="""
You are an expert research assistant.
You are provided with a Context in JSON format and a Question.
Use RAG to answer the Question, providing references and links to the Context material you retrieve and use in your answer:
When generating your answer, follow these steps:
- Retrieve the most relevant context material from your knowledge base to help answer the question
- Cite the references you use by including the title, author, publication, and a link to each source
- Synthesize the retrieved information into a clear, informative answer to the question
- Format your answer in Markdown, using heading levels 2-3 as needed
- Include a "References" section at the end with the full citations and link for each source you used
Example of Context JSON entry:
{
"page_content": "This provides access to material related to ...",
"metadata": {
"title": "Introduction - Marie Curie: Topics in Chronicling America",
"link": "https://guides.loc.gov/chronicling-america-marie-curie"
}
}
"""
),
HumanMessage(
content= f"""
Context information is below.
Context:
---------------------
{json.dumps(relevant_docs, indent=2, ensure_ascii=False)}
---------------------
Question: {query}
Answer:
"""
),
]
response = chat.invoke(messages, config={"callbacks": callbacks})
return response
console = Console()
dotenv.load_dotenv()
callbacks = []
if(os.getenv("LANGCHAIN_API_KEY")):
callbacks.append(
LangChainTracer(
project_name="search agent",
client=Client(
api_url="https://api.smith.langchain.com",
)
)
)
if __name__ == '__main__':
arguments = docopt(__doc__, version='Search Agent 0.1')
#print(arguments)
provider = arguments["--provider"]
temperature = float(arguments["--temperature"])
chat = get_chat_llm(provider, temperature)
query = arguments["SEARCH_QUERY"]
with console.status(f"[bold green]Optimizing query for search: {query}"):
optimize_search_query = optimize_search_query(query)
console.log(f"Optimized search query: [bold blue]{optimize_search_query}")
domain=arguments["--domain"]
max_pages=arguments["--max_pages"]
with console.status(f"[bold green]Searching sources using the optimized query: {optimize_search_query}"):
sources = get_sources(optimize_search_query, max_pages=max_pages, domain=domain)
console.log(f"Found {len(sources)} sources {'on ' + domain if domain else ''}")
with console.status(f"[bold green]Fetching content for {len(sources)} sources", spinner="growVertical"):
contents = get_links_contents(sources)
console.log(f"Managed to extract content from {len(contents)} sources")
with console.status(
f"[bold green]Processing {len(contents)} contents and finding relevant extracts",
spinner="dots8Bit"
):
relevant_docs = process_and_vectorize_content(contents, query)
console.log(f"Filtered {len(relevant_docs)} relevant content extracts")
with console.status(f"[bold green]Querying LLM with {len(relevant_docs)} relevant extracts", spinner='dots8Bit'):
respomse = answer_query_with_sources(query, relevant_docs)
console.rule(f"[bold green]Response from {provider}")
console.print(Markdown(respomse.content))
console.rule("[bold green]")