app.py

import gradio as gr
import sys
from utils.retriever import get_context, get_vectorstore

# Initialize vector store at startup
print("Initializing vector store connection...", flush=True)
try:
    vectorstore = get_vectorstore()
    print("Vector store connection initialized successfully", flush=True)
except Exception as e:
    print(f"Failed to initialize vector store: {e}", flush=True)
    raise

# ---------------------------------------------------------------------
# MCP - returns raw dictionary format
# ---------------------------------------------------------------------

def retrieve(
    query:str, 
    collection_name:str = None,
    filter_metadata:dict = None,
) -> list:
    """
    Retrieve semantically similar documents from the vector database for MCP clients.
    
    Args:
        query (str): The search query text
        reports_filter (str): Comma-separated list of specific report filenames (optional)
        sources_filter (str): Filter by document source type (optional)  
        subtype_filter (str): Filter by document subtype (optional)
        year_filter (str): Comma-separated list of years to filter by (optional)
    
    Returns:
        list: List of dictionaries containing document content, metadata, and scores
    """

    
    # Call retriever function and return raw results
    results = get_context(
        vectorstore=vectorstore,
        query=query,
        collection_name=collection_name,
        filter_metadata = filter_metadata
    )
    
    return results


# Create the Gradio interface with Blocks to support both UI and MCP
with gr.Blocks() as ui:
    gr.Markdown("# ChatFed Retrieval/Reranker Module")
    gr.Markdown("Retrieves semantically similar documents from vector database and reranks. Intended for use in RAG pipelines as an MCP server with other ChatFed modules.")
    
    with gr.Row():
        with gr.Column():
            query_input = gr.Textbox(
                label="Query", 
                lines=2, 
                placeholder="Enter your search query here",
                info="The query to search for in the vector database"
            )
            collection_name = gr.Textbox(
                label="Collection Name (optional)", 
                lines=1, 
                placeholder="EUDR, Humboldt",
                info="Name of the collection"
            )
            sources_input = gr.Textbox(
                label="Sources Filter key to be looked in metadata (optional)", 
                lines=1, 
                placeholder="country",
                info="Filter by document source type (leave empty for all)"
            )
            sources_value = gr.Textbox(
                label="Value in filter to be looked for(optional)", 
                lines=1, 
                placeholder="Ecuador, Guatemala",
                info="Filter by document subtype (leave empty for all)"
            )
            if sources_input & sources_value:
                filter_metadata = {sources_input:sources_value}
            else: 
                filter_metadata = None
            
            submit_btn = gr.Button("Submit", variant="primary")
        
        # Output needs to be in json format to be added as tool in HuggingChat
        with gr.Column():
            output = gr.Text(
                label="Retrieved Context",
                lines=10,
                show_copy_button=True
            )
    
    # UI event handler
    submit_btn.click(
        fn=retrieve,
        inputs=[query_input,collection_name, filter_metadata],
        outputs=output,
        api_name="retrieve"
    )
    

# Launch with MCP server enabled
if __name__ == "__main__":
    ui.launch(
        server_name="0.0.0.0",
        server_port=7860, 
        #mcp_server=True,
        show_error=True
    )