You run an MCP server that combines a contextual ML FAQ retriever, web search via Bright Data proxy, and structured prompts for Claude Desktop. It lets you answer ML questions from a vector FAQ, perform web research through a premium proxy, and interact with Claude Desktop using predefined prompts, all served through a single MCP endpoint for your client applications.

How to use

You can operate the MCP server from your local machine and connect your MCP client to the local endpoint. Use the built-in tools to (1) retrieve answers from a machine learning FAQ base, (2) perform web searches through Bright Data, and (3) interact with Claude Desktop using structured prompts.

To answer ML questions from the FAQ, invoke the machine_learning_faq_retrieval_tool with a question such as:

Use a ferramenta "machine_learning_faq_retrieval_tool" para responder perguntas como:
"Como evitar overfitting em modelos de machine learning?"```For general web research via Bright Data, use the bright_data_web_search_tool with a query like:```Você é um assistente de conhecimento geral.
Use "bright_data_web_search_tool" para perguntas como:
"Qual o impacto atual da inteligência artificial no setor bancário brasileiro?"```
These prompts guide the MCP client to route the user query to the corresponding tool and compose an answer from the retrieved data or search results.

## How to install

Prerequisites: Python 3.11+ installed on your system. You will also need Docker if you plan to run Qdrant locally.

git clone https://seurepo.com/mcp_agentic_rag.git cd mcp_agentic_rag python -m venv .venv .venv\Scripts\activate pip install -r requirements.txt

Bright Data credentials will be stored in the environment file

Create the Bright Data credential file as described in the next step

Create credentials for Bright Data in a file named ".env" with your Bright Data username and password.

BRIGHT_DATA_USERNAME=brd-customer-usuario-zone-nomezona BRIGHT_DATA_PASSWORD=sua_senha

Start Qdrant in a local container to host the FAQ embeddings.

docker run -p 6333:6333 -p 6334:6334
-v ${PWD}/qdrant_storage:/qdrant/storage
qdrant/qdrant

Ingest the FAQ into the vector store and prepare embeddings.

python ingest_faq.py

You should see a success message indicating the collection was populated in Qdrant.

Launch the MCP server that ties everything together.

python server.py

## Additional configuration and notes

Configure Claude Desktop to connect to your local MCP server by adding a new entry under Settings > MCP Servers with the following details. This example assumes your server runs locally on port 8080.

{ "mcpServers": { "mcp_rag_app": { "command": "python", "args": ["server.py"], "host": "127.0.0.1", "port": 8080, "timeout": 30000 } } }

## Security and maintenance notes

Keep your Bright Data credentials secure and do not commit the ".env" file to version control. Update embeddings and re-ingest FAQs periodically to keep the vector store fresh. Monitor the server logs for any connection issues with Qdrant or the Claude Desktop client.

## Available tools

### machine\_learning\_faq\_retrieval\_tool

Retrieves answers from a machine-learning FAQ vector store to respond ML-related questions.

### bright\_data\_web\_search\_tool

Performs web searches using the Bright Data proxy for up-to-date information.

### claude\_desktop\_client

Interacts with Claude Desktop using structured prompts to obtain responses.