LangChain: Add example using MCP

.github/workflows/ml-langchain.yml

@@ -38,7 +38,12 @@ jobs:
       matrix:
         os: [ 'ubuntu-latest' ]
         python-version: [ '3.10', '3.11', '3.12', '3.13' ]
-        cratedb-version: [ 'nightly' ]
+        cratedb-version: [
+          'nightly',
+        ]
+        cratedb-mcp-version: [
+          'pr-50',
+        ]
 
     services:
       cratedb:
@@ -48,6 +53,15 @@ jobs:
           - 5432:5432
         env:
           CRATE_HEAP_SIZE: 4g
+      cratedb-mcp:
+        image: ghcr.io/crate/cratedb-mcp:${{ matrix.cratedb-mcp-version }}
+        ports:
+          - 8000:8000
+        env:
+          CRATEDB_MCP_TRANSPORT: streamable-http
+          CRATEDB_MCP_HOST: 0.0.0.0
+          CRATEDB_MCP_PORT: 8000
+          CRATEDB_CLUSTER_URL: http://crate:crate@cratedb:4200/
 
     env:
       OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

topic/machine-learning/llm-langchain/.gitignore

@@ -1,2 +1,3 @@
-*.sql
 .env
+*.sql
+!init.sql

topic/machine-learning/llm-langchain/README.md

@@ -7,13 +7,18 @@
 by language models. It provides a complete set of powerful and flexible
 components for building context-aware, reasoning applications.
 
-Please refer to the [LangChain documentation] for further information.
+[LangGraph] is a low-level orchestration framework for building, managing,
+and deploying long-running, stateful agents.
+
+Please refer to the [LangChain documentation] and the [Building Ambient
+Agents with LangGraph] academy material for further information.
 
 Common end-to-end use cases are:
 
 - Analyzing structured data
 - Chatbots and friends
 - Document question answering
+- Text-to-SQL (talk to your data)
 
 LangChain provides standard, extendable interfaces and external integrations
 for the following modules, listed from least to most complex:
@@ -79,17 +84,21 @@ and [CrateDB].
   augmented generation (RAG) pipeline. To implement RAG we use the Python client driver for 
   CrateDB and vector store support in LangChain.
 
-  - `cratedb_rag_customer_support_vertexai.ipynb` [![Open on GitHub](https://img.shields.io/badge/Open%20on-GitHub-lightgray?logo=GitHub)](cratedb_rag_customer_support_vertexai.ipynb)[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/crate/cratedb-examples/blob/main/topic/machine-learning/llm-langchain/cratedb_rag_customer_support_vertexai.ipynb)
+- `cratedb_rag_customer_support_vertexai.ipynb` [![Open on GitHub](https://img.shields.io/badge/Open%20on-GitHub-lightgray?logo=GitHub)](cratedb_rag_customer_support_vertexai.ipynb)[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/crate/cratedb-examples/blob/main/topic/machine-learning/llm-langchain/cratedb_rag_customer_support_vertexai.ipynb)
 
   This example illustrates the RAG implementation of a customer support scenario. 
   It is based on the previous notebook, and it illustrates how to use Vertex AI platform
   on Google Cloud for RAG pipeline.
-
 
+- `agent_with_mcp.py`
+
+  This example illustrates how to use LangGraph and the `langchain-mcp-adapters`
+  package to implement an LLM agent that is connecting to the CrateDB MCP server.
+  The demo program performs Text-to-SQL on timeseries data stored in a CrateDB table.
 
 ## Install
 
-In order to properly set up a sandbox environment to explore the example notebooks
+To properly set up a sandbox environment to explore the example notebooks
 and programs, it is advised to create a Python virtualenv, and install the
 dependencies into it. In this way, it is easy to wipe your virtualenv and start
 from scratch anytime.
@@ -126,6 +135,26 @@ a cloud-based development environment is up and running. As soon as your project
 easily move to a different cluster tier or scale horizontally.
 
 
+### MCP
+
+Spin up the [CrateDB MCP server], connecting it to CrateDB on localhost.
+```bash
+export CRATEDB_CLUSTER_URL=http://crate:crate@localhost:4200/
+export CRATEDB_MCP_TRANSPORT=streamable-http
+uvx cratedb-mcp serve
+```
+
+Run the code using OpenAI API:
+```bash
+export OPENAI_API_KEY=<YOUR_OPENAI_API_KEY>
+python agent_with_mcp.py
+```
+Expected output:
+```text
+Query was: What is the average value for sensor 1?
+Answer was: The average value for sensor 1 is approximately 17.03. If you need more details or a different calculation, let me know!
+```
+
 ## Testing
 
 Run all tests.
@@ -139,22 +168,25 @@ pytest -k document_loader
 pytest -k "notebook and loader"
 ```
 
-In order to force a regeneration of the Jupyter Notebook, use the
+To force a regeneration of the Jupyter Notebook, use the
 `--nb-force-regen` option.
 ```shell
 pytest -k document_loader --nb-force-regen
 ```
 
 
 [Agents]: https://python.langchain.com/docs/modules/agents/
+[Building Ambient Agents with LangGraph]: https://academy.langchain.com/courses/ambient-agents/
 [Callbacks]: https://python.langchain.com/docs/modules/callbacks/
 [Chains]: https://python.langchain.com/docs/modules/chains/
 [CrateDB]: https://github.com/crate/crate
 [CrateDB Cloud]: https://console.cratedb.cloud
+[CrateDB MCP server]: https://cratedb.com/docs/guide/integrate/mcp/cratedb-mcp.html
 [`FLOAT_VECTOR`]: https://crate.io/docs/crate/reference/en/master/general/ddl/data-types.html#float-vector
 [`KNN_MATCH`]: https://crate.io/docs/crate/reference/en/master/general/builtins/scalar-functions.html#scalar-knn-match
 [LangChain]: https://www.langchain.com/
 [LangChain documentation]: https://python.langchain.com/
+[LangGraph]: https://langchain-ai.github.io/langgraph/
 [Memory]: https://python.langchain.com/docs/modules/memory/
 [Model I/O]: https://python.langchain.com/docs/modules/model_io/
 [Retrieval]: https://python.langchain.com/docs/modules/data_connection/

topic/machine-learning/llm-langchain/agent_with_mcp.py

@@ -0,0 +1,54 @@
+"""
+Exercise LangChain/LangGraph with the CrateDB MCP server.
+
+## Synopsis
+
+# Install prerequisites.
+pip install -U -r requirements.txt
+
+# Start database.
+docker run --rm -it --publish=4200:4200 crate/crate:nightly
+
+# Start MCP server.
+export CRATEDB_MCP_TRANSPORT=streamable-http
+export CRATEDB_MCP_HOST=0.0.0.0
+export CRATEDB_MCP_PORT=8000
+export CRATEDB_CLUSTER_URL=http://crate:crate@localhost:4200/
+docker run --rm -it --network=host --publish=8000:8000 ghcr.io/crate/cratedb-mcp:pr-50
+
+# Run program.
+export OPENAI_API_KEY=<your_openai_api_key>
+python agent_with_mcp.py
+"""
+import asyncio
+
+from langchain_mcp_adapters.client import MultiServerMCPClient
+from langgraph.prebuilt import create_react_agent
+
+
+async def amain():
+    client = MultiServerMCPClient(
+        {
+            "cratedb": {
+                "transport": "streamable_http",
+                "url": "http://localhost:8000/mcp/"
+            },
+        }
+    )
+    tools = await client.get_tools()
+    agent = create_react_agent("openai:gpt-4.1", tools)
+
+    QUERY_STR = "What is the average value for sensor 1?"
+    response = await agent.ainvoke({"messages": QUERY_STR})
+    answer = response["messages"][-1].content
+
+    print("Query was:", QUERY_STR)
+    print("Answer was:", answer)
+
+
+def main():
+    asyncio.run(amain())
+
+
+if __name__ == "__main__":
+    main()

topic/machine-learning/llm-langchain/init.sql

@@ -0,0 +1,25 @@
+DROP TABLE IF EXISTS time_series_data;
+
+CREATE TABLE IF NOT EXISTS time_series_data (
+    timestamp TIMESTAMP,
+    value DOUBLE,
+    location STRING,
+    sensor_id INT
+);
+
+INSERT INTO time_series_data (timestamp, value, location, sensor_id)
+VALUES
+    ('2023-09-14T00:00:00', 10.5, 'Sensor A', 1),
+    ('2023-09-14T01:00:00', 15.2, 'Sensor A', 1),
+    ('2023-09-14T02:00:00', 18.9, 'Sensor A', 1),
+    ('2023-09-14T03:00:00', 12.7, 'Sensor B', 2),
+    ('2023-09-14T04:00:00', 17.3, 'Sensor B', 2),
+    ('2023-09-14T05:00:00', 20.1, 'Sensor B', 2),
+    ('2023-09-14T06:00:00', 22.5, 'Sensor A', 1),
+    ('2023-09-14T07:00:00', 18.3, 'Sensor A', 1),
+    ('2023-09-14T08:00:00', 16.8, 'Sensor A', 1),
+    ('2023-09-14T09:00:00', 14.6, 'Sensor B', 2),
+    ('2023-09-14T10:00:00', 13.2, 'Sensor B', 2),
+    ('2023-09-14T11:00:00', 11.7, 'Sensor B', 2);
+
+REFRESH TABLE time_series_data;

topic/machine-learning/llm-langchain/requirements.txt

@@ -1,9 +1,11 @@
 crash
 google-cloud-aiplatform<2
 langchain-cratedb<0.1.2
+langchain-mcp-adapters<0.2
 langchain-google-vertexai<3
 langchain-openai<0.4
 langchain-text-splitters<0.4
+langgraph<0.6
 pueblo[cli,nlp]>=0.0.10
 pypdf<6
 python-dotenv<2

topic/machine-learning/llm-langchain/test.py

@@ -29,6 +29,14 @@ def reset_database(cratedb):
     time.sleep(0.01)
 
 
+@pytest.fixture(scope="function", autouse=True)
+def init_database(cratedb):
+    """
+    Initialize database.
+    """
+    cratedb.run_sql((HERE / "init.sql").read_text())
+
+
 def pytest_generate_tests(metafunc):
     """
     Generate pytest test case per Jupyter Notebook.