feat(MCP): Make Wiki-RAG to behave as a MCP server

README.md

@@ -59,6 +59,7 @@ To get started with Wiki-RAG, ensure you have the following:
    - `wr-index`: In charge of creating the collection in the vector index (Milvus) with all the information extracted in the previous step.
    - `wr-search`: A tiny CLI utility to perform searches against the RAG system from the command line.
    - `wr-server`: A comprehensive and secure web service (documented with OpenAPI) that allows users to interact with the RAG system using the OpenAI API (`v1/models` and `v1/chat/completions` endpoints) as if it were a large language model (LLM).
+   - `wr-mcp`: A complete and **UNPROTECTED** built-in MCP server that allows you to access to various parts of Wiki-RAG like prompts (system and use prompt with placeholders), resources (access to the source parsed documents) and tools (retrieve, optimise and generate) using the [MCP Protocol](https://modelcontextprotocol.io/).
 
 ### Running with Docker (Milvus elsewhere)
 

env_template

@@ -32,6 +32,9 @@ WRAPPER_CHAT_MAX_TOKENS=1536
 # Public name that the wrapper will use to identify itself as a model. Will default to COLLECTION_NAME is not set.
 WRAPPER_MODEL_NAME="Your Model Name"
 
+# Model Context Protocol (MCP) base URL.
+MCP_API_BASE="0.0.0.0:8081"
+
 # Validate requests auth against these bearer tokens.
 AUTH_TOKENS="11111111,22222222,33333333"
 # Delegate bearer token auth to this URL.

pyproject.toml

@@ -36,6 +36,7 @@ dependencies = [
     "langchain-openai == 0.3.7",
     "langgraph == 0.3.5",
     "langsmith == 0.3.11",
+    "mcp[cli] == 1.5.0",
     "pymilvus == 2.5.4",
     "python-dotenv == 1.0.1",
     "tiktoken == 0.9.0",
@@ -60,6 +61,7 @@ wr-load = "wiki_rag.load.main:main"
 wr-index = "wiki_rag.index.main:main"
 wr-search = "wiki_rag.search.main:main"
 wr-server = "wiki_rag.server.main:main"
+wr-mcp = "wiki_rag.mcp_server.main:main"
 
 [project.urls]
 Homepage = "https://github.com/moodlehq/wiki-rag"

wiki_rag/mcp_server/__init__.py

@@ -0,0 +1,9 @@
+#  Copyright (c) 2025, Moodle HQ - Research
+#  SPDX-License-Identifier: BSD-3-Clause
+
+"""wiki_rag.mcp_server package."""
+
+from pathlib import Path
+
+# The file that will be used to provide resources.
+res_file: Path | None = None

wiki_rag/mcp_server/main.py

@@ -0,0 +1,170 @@
+#  Copyright (c) 2025, Moodle HQ - Research
+#  SPDX-License-Identifier: BSD-3-Clause
+
+"""Main entry point for the knowledge base MCP compatible server."""
+
+import logging
+import os
+import sys
+
+from pathlib import Path
+
+from dotenv import load_dotenv
+from langchain_core.runnables import RunnableConfig
+
+import wiki_rag.index as index
+import wiki_rag.mcp_server as mcp_global
+
+from wiki_rag import LOG_LEVEL, ROOT_DIR, __version__, server
+from wiki_rag.search.util import ConfigSchema
+from wiki_rag.util import setup_logging
+
+
+def main():
+    """Run the MCP server with all the configuration in place."""
+    setup_logging(level=LOG_LEVEL)
+    logger = logging.getLogger(__name__)
+    logger.info("wiki_rag-server-mcp_server starting up...")
+
+    # Print the version of the bot.
+    logger.warning(f"Version: {__version__}")
+
+    dotenv_file = ROOT_DIR / ".env"
+    if dotenv_file.exists():
+        logger.warning("Loading environment variables from %s", dotenv_file)
+        logger.warning("Note: .env files are not supposed to be used in production. Use env secrets instead.")
+        load_dotenv(dotenv_file)
+
+    mediawiki_url = os.getenv("MEDIAWIKI_URL")
+    if not mediawiki_url:
+        logger.error("Mediawiki URL not found in environment. Exiting.")
+        sys.exit(1)
+
+    mediawiki_namespaces = os.getenv("MEDIAWIKI_NAMESPACES")
+    if not mediawiki_namespaces:
+        logger.error("Mediawiki namespaces not found in environment. Exiting.")
+        sys.exit(1)
+    mediawiki_namespaces = mediawiki_namespaces.split(",")
+    mediawiki_namespaces = [int(ns.strip()) for ns in mediawiki_namespaces]  # no whitespace and int.
+    mediawiki_namespaces = list(set(mediawiki_namespaces))  # unique
+
+    loader_dump_path = os.getenv("LOADER_DUMP_PATH")
+    if loader_dump_path:
+        loader_dump_path = Path(loader_dump_path)
+    else:
+        loader_dump_path = ROOT_DIR / "data"
+    # If the directory does not exist, create it.
+    if not loader_dump_path.exists():
+        logger.warning(f"Data directory {loader_dump_path} not found. Creating it.")
+        try:
+            loader_dump_path.mkdir()
+        except Exception:
+            logger.error(f"Could not create data directory {loader_dump_path}. Exiting.")
+            sys.exit(1)
+
+    collection_name = os.getenv("COLLECTION_NAME")
+    if not collection_name:
+        logger.error("Collection name not found in environment. Exiting.")
+        sys.exit(1)
+        # TODO: Validate that only numbers, letters and underscores are used.
+
+    index.milvus_url = os.getenv("MILVUS_URL")
+    if not index.milvus_url:
+        logger.error("Milvus URL not found in environment. Exiting.")
+        sys.exit(1)
+
+    # If tracing is enabled, put a name for the project.
+    if os.getenv("LANGSMITH_TRACING", "false") == "true":
+        os.environ["LANGSMITH_PROJECT"] = f"{collection_name}"
+
+    user_agent = os.getenv("USER_AGENT")
+    if not user_agent:
+        logger.info("User agent not found in environment. Using default.")
+        user_agent = "Moodle Research Crawler/{version} (https://git.in.moodle.com/research)"
+    user_agent = f"{user_agent.format(version=__version__)}"
+
+    embedding_model = os.getenv("EMBEDDING_MODEL")
+    if not embedding_model:
+        logger.error("Embedding model not found in environment. Exiting.")
+        sys.exit(1)
+
+    embedding_dimensions = os.getenv("EMBEDDING_DIMENSIONS")
+    if not embedding_dimensions:
+        logger.error("Embedding dimensions not found in environment. Exiting.")
+        sys.exit(1)
+    embedding_dimensions = int(embedding_dimensions)
+
+    llm_model = os.getenv("LLM_MODEL")
+    if not llm_model:
+        logger.error("LLM model not found in environment. Exiting.")
+        sys.exit(1)
+
+    mcp_api_base = os.getenv("MCP_API_BASE")
+    if not mcp_api_base:
+        logger.error("MCP API base not found in environment. Exiting.")
+        sys.exit(1)
+    parts = mcp_api_base.split(":")
+    mcp_server = parts[0]
+    if len(parts) > 1:
+        mcp_port = int(parts[1])
+    else:
+        mcp_port = 8081
+
+    # Calculate the file that we are going to use as source for the resources.
+    input_candidate = ""
+    for file in sorted(loader_dump_path.iterdir()):
+        if file.is_file() and file.name.startswith(collection_name) and file.name.endswith(".json"):
+            input_candidate = file
+    if input_candidate:
+        mcp_global.res_file = loader_dump_path / input_candidate
+
+    if not mcp_global.res_file:
+        logger.warning(f"No input file found in {loader_dump_path} with collection name {collection_name}.")
+
+    # These are optional, default to 0 (unlimited).
+    wrapper_chat_max_turns = int(os.getenv("WRAPPER_CHAT_MAX_TURNS", 0))
+    wrapper_chat_max_tokens = int(os.getenv("WRAPPER_CHAT_MAX_TOKENS", 0))
+    wrapper_model_name = os.getenv("WRAPPER_MODEL_NAME") or os.getenv("COLLECTION_NAME")
+    if not wrapper_model_name:
+        logger.error("Public wrapper name not found in environment. Exiting.")  # This is unreachable.
+        sys.exit(1)
+
+    # Prepare the configuration schema.
+    # TODO, make prompt name, task_def, kb_*, cutoff, max tokens, temperature, top_p
+    #  configurable. With defaults applied if not configured.
+    config_schema = ConfigSchema(
+        prompt_name="moodlehq/wiki-rag",
+        task_def="Moodle user documentation",
+        kb_name="Moodle Docs",
+        kb_url=mediawiki_url,
+        collection_name=collection_name,
+        embedding_model=embedding_model,
+        embedding_dimension=embedding_dimensions,
+        llm_model=llm_model,
+        search_distance_cutoff=0.6,
+        max_completion_tokens=768,
+        temperature=0.05,
+        top_p=0.85,
+        stream=False,
+        wrapper_chat_max_turns=wrapper_chat_max_turns,
+        wrapper_chat_max_tokens=wrapper_chat_max_tokens,
+        wrapper_model_name=wrapper_model_name,
+    ).items()
+
+    # Prepare the configuration.
+    server.config = RunnableConfig(configurable=dict(config_schema))
+
+    # Start the mcp_server server
+    from wiki_rag.mcp_server.server import mcp
+
+    mcp.settings.host = mcp_server
+    mcp.settings.port = mcp_port
+    mcp.run("sse")
+    # import asyncio
+    # asyncio.run(mcp_server.run_sse_async())
+
+    logger.info("wiki_rag-server-mcp_server finished.")
+
+
+if __name__ == "__main__":
+    main()