kb-mcp-server

# QA MCP Test This directory contains scripts for testing the txtai MCP server with a question-answering dataset. ## Files - `qa.yml` - Configuration file for the txtai embeddings - `load_web_questions.py` - Script to load the web_questions dataset from Hugging Face into txtai embeddings - `test_qa_server.py` - Script to test the QA functionality with the MCP server ## Usage ### 1. Load the web_questions dataset First, load the web_questions dataset into txtai embeddings: ```bash # Install required dependencies pip install datasets txtai # Load the full dataset python load_web_questions.py --test # Or load a limited number of questions (e.g., 100) python load_web_questions.py --limit 100 --test # Save the embeddings as a tar.gz archive python load_web_questions.py --limit 100 --save-archive qa_embeddings.tar.gz ``` This will: - Download the web_questions dataset from Hugging Face - Create a txtai embeddings index in `.txtai/indexes/qa` - Index the questions and answers - Run test queries if `--test` is specified - Save the embeddings as a tar.gz archive if `--save-archive` is specified ### 2. Test the QA functionality with the MCP server After loading the dataset, test that the QA functionality works with the MCP server: ```bash # Test using the configuration file (default) python test_qa_server.py # Test using a directory of embeddings python test_qa_server.py --embeddings .txtai/indexes/qa # Test using an archive file python test_qa_server.py --embeddings qa_embeddings.tar.gz ``` This will: - Load the embeddings from the specified source (config file, directory, or archive) - Test the answer_question functionality directly ### 3. Run the MCP server with the QA embeddings Finally, run the MCP server with the QA embeddings: ```bash cd ../../ # Using a directory mcp run --transport sse server.py --embeddings .txtai/indexes/qa # Or using an archive file mcp run --transport sse server.py --embeddings test/qa-mcp/qa_embeddings.tar.gz ``` This will: - Start the MCP server - Load the QA embeddings (from either a directory or an archive file) - Register the search and QA tools ## Notes - The web_questions dataset contains about 3,778 question-answer pairs - The embeddings are stored using SQLite for persistence - The extractor pipeline uses the `distilbert-base-cased-distilled-squad` model for extractive QA - When creating a tar.gz archive, always use the `--save-archive` option or the `embeddings.save()` method, as manually archiving the directory will not create the correct internal structure