Use-case

Approach

**IterDRAG Inspiration**

• By iteratively retrieving documents and generating answers
• Continuously generates new research focuses based on what it discovers
• Gathers content from each search, saving that content into state w/ links

Specific Approach

Local LLM

Open LLM leaderboard

huggingface.co

Ollama

• Can also try qwen2.5 etc

ollama pull llama3.2

Code

from langsmith import traceable

def deduplicate_and_format_sources(search_response, max_tokens_per_source, include_raw_content=True):
    """
    Takes either a single search response or list of responses from Tavily API and formats them.
    Limits the raw_content to approximately max_tokens_per_source.
    include_raw_content specifies whether to include the raw_content from Tavily in the formatted string.
    
    Args:
        search_response: Either:
            - A dict with a 'results' key containing a list of search results
            - A list of dicts, each containing search results
            
    Returns:
        str: Formatted string with deduplicated sources
    """
    # Convert input to list of results
    if isinstance(search_response, dict):
        sources_list = search_response['results']
    elif isinstance(search_response, list):
        sources_list = []
        for response in search_response:
            if isinstance(response, dict) and 'results' in response:
                sources_list.extend(response['results'])
            else:
                sources_list.extend(response)
    else:
        raise ValueError("Input must be either a dict with 'results' or a list of search results")
    
    # Deduplicate by URL
    unique_sources = {}
    for source in sources_list:
        if source['url'] not in unique_sources:
            unique_sources[source['url']] = source
    
    # Format output
    formatted_text = "Sources:\\n\\n"
    for i, source in enumerate(unique_sources.values(), 1):
        formatted_text += f"Source {source['title']}:\\n===\\n"
        formatted_text += f"URL: {source['url']}\\n===\\n"
        formatted_text += f"Most relevant content from source: {source['content']}\\n===\\n"
        if include_raw_content:
            # Using rough estimate of 4 characters per token
            char_limit = max_tokens_per_source * 4
            # Handle None raw_content
            raw_content = source.get('raw_content', '')
            if raw_content is None:
                raw_content = ''
                print(f"Warning: No raw_content found for source {source['url']}")
            if len(raw_content) > char_limit:
                raw_content = raw_content[:char_limit] + "... [truncated]"
            formatted_text += f"Full source content limited to {max_tokens_per_source} tokens: {raw_content}\\n\\n"
                
    return formatted_text.strip()

def format_sources(search_results):
    """Format search results into a bullet-point list of sources.
    
    Args:
        search_results (dict): Tavily search response containing results
        
    Returns:
        str: Formatted string with sources and their URLs
    """
    return '\\n'.join(
        f"* {source['title']} : {source['url']}"
        for source in search_results['results']
    )

@traceable
def tavily_search(query, include_raw_content=True, max_results=3):
    """ Search the web using the Tavily API.
    
    Args:
        query (str): The search query to execute
        include_raw_content (bool): Whether to include the raw_content from Tavily in the formatted string
        max_results (int): Maximum number of results to return
        
    Returns:
        dict: Tavily search response containing:
            - results (list): List of search result dictionaries, each containing:
                - title (str): Title of the search result
                - url (str): URL of the search result
                - content (str): Snippet/summary of the content
                - raw_content (str): Full content of the page if available"""
     
    return tavily_client.search(query, 
                         max_results=max_results, 
                         include_raw_content=include_raw_content)

from langchain_ollama import ChatOllama

### LLM
local_llm = "llama3.2"
local_llm = "qwen2.5:14b"
llm = ChatOllama(model=local_llm, temperature=0)
llm_json_mode = ChatOllama(model=local_llm, temperature=0, format="json")

### Search
from tavily import TavilyClient
tavily_client = TavilyClient()