Add ReAct agent and streamlit web app

config.yaml

@@ -0,0 +1,23 @@
+tools:
+  - name: "constitution_tool"
+    description: "Answers questions about the U.S. Constitution."
+    url: "https://my.app/v1/completions"
+    config:
+      method: 'POST'
+      headers:
+        'Content-Type': 'application/json'
+        'Authorization': 'Basic 12345'
+      body:
+        prompt: '{{prompt}}'
+      responseParser: 'json.answer'
+      responseMetadata: 
+        - name: 'sources'
+          loc: 'json.sources'
+      responseFormat:
+        agent: '{{response}}'
+        json: 
+          - "response"
+          - "sources"
+    examples:
+      - "What is the definition of a citizen in the U.S. Constitution?"
+      - "What article describes the power of the judiciary branch?"

pyproject.toml

@@ -0,0 +1,53 @@
+[tool.poetry]
+name = "llm_agents"
+version = "0.1.0"
+description = ""
+readme = "README.md"
+authors = ["Red Hat"]
+packages = [{include = "llm_agents"}]
+
+[tool.poetry.dependencies]
+python = ">=3.10,<4.0"
+pyarrow = "15.0.0"
+langchain = "^0.2.5"
+langchain-openai = "^0.1.9"
+langchain-community = "^0.2.5"
+mlflow = "^2.14.1"
+python-dotenv = "^1.0.1"
+fastapi = "^0.111.0"
+
+[tool.poetry.group.dev.dependencies]
+pytest = "^8.1"
+pytest-cov = "^4.0.0"
+ruff = "^0.4.1"
+bump2version = "^1.0.1"
+
+[tool.poetry.group.webapp.dependencies]
+streamlit = "~1.32.0"
+
+[[tool.poetry.source]]
+name = "PyPI"
+priority = "primary"
+
+[[tool.poetry.source]]
+name = "de-cop-nexus"
+url = "https://nexus.corp.redhat.com/repository/de-cop-pypi-releases/simple/"
+priority = "supplemental"
+
+[tool.ruff]
+line-length = 140
+
+[tool.ruff.lint]
+select = ["E", "F", "D", "C", "N"]
+ignore = [
+    "E501", # line-too-long
+    "E402", # module-import-not-at-top-file
+    "D203", # one-blank-line-before-class
+    "D212", # multi-line-summary-first-line
+    "D100", # undocumented-public-module
+]
+per-file-ignores = { "tests/*" = ["D"] }
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"

react_agent/README.md

@@ -0,0 +1,123 @@
+# Instructions
+
+## Pre-Requisites
+
+In order to run the ReAct agent setup, the following tools must be installed:
+* `poetry`
+* `oc` (optional)
+* [Ollama](https://ollama.com/)
+
+**Initial Steps**
+
+* Clone the repository to your local system via `git clone [email protected]:redhat-et/llm-agents.git`
+
+* Change directory to the project `cd llm-agents`
+
+* Install the project dependencies `poetry install`
+
+## Local Usage
+
+To run the setup locally, follow the steps below:
+
+1. Create a copy of the `sample.env` and rename it as `.env`, then fill in any environment variables
+    * The `sample.env` is configured to use Ollama by default, so if you are using that as your LLM provider you don't need to change anything after copying. If you want to use any other LLM then you will need to update `OPENAI_URI` with the model serving endpoint.
+
+2. The `config.yaml` provided in the repo has an example for a dummy `Constitution Tool`. Update the `config.yaml` with information for any tool/API endpoints you'd like the ReAct agent to interact with.
+
+***NOTE**: If you do not wish to use any APIs, remove the "tools" header completely.*
+
+3. If using Ollama, ensure that the model specified in `OPENAI_MODEL` has been pulled into your Ollama instance with `ollama pull <model>`. You can view all the models available in your Ollama instance by running `ollama list` and ensure that the model name in `OPENAI_MODEL` is the same as the ones listed. 
+
+4. Run `ollama serve`. If you get an error like "Port 11434 already in use" then Ollama is running and you can move to the next step. 
+
+5. In a new terminal, run `poetry run mlflow server` to spin up the MLFlow tracking server. This will run a local instance of MLFlow on your system to log the traces/outputs of the agent application. Navigate to the URL provided to view the MLFlow UI and you will see the outputs being logged under the experiment name `ReAct Agent` -> click on the `Traces` tab.
+
+6. In a new terminal, run `poetry run python react_agent/api.py` to spin up the agent API server.
+
+7. In a new terminal, run `poetry run streamlit run streamlit/intro.py` and navigate to the URL provided to use the UI application. You can now chat with the application by asking a question and see the outputs being generated.
+
+## Adding Tools for the ReAct agent
+
+There are two ways you can add tools to the ReAct agent's toolbelt. 
+Please note that for any of these tools, currently only a single input and a single output is supported.
+We may add support for structured inputs in the future.
+
+### Option 1: Custom Python Tool (Easiest)
+To create a custom Python tool, open `react_agent/tools`.
+Here you will see a few different categories of tools already defined.
+In the `math.py` file, you'll see the most basic definition of a tool called `ComputeSquareTool`, which computes the square of a number.
+You can use this tool as an example for your tool.
+
+```python
+from typing import Optional, Type
+
+from langchain.callbacks.manager import (
+    AsyncCallbackManagerForToolRun,
+    CallbackManagerForToolRun,
+)
+from langchain_core.tools import BaseTool
+from pydantic import BaseModel, Field
+
+class ComputeSquareInput(BaseModel):
+    """Compute square tool input structure."""
+
+    number: int = Field(description="number to square")
+
+
+class ComputeSquareTool(BaseTool):
+    """Tool for computing the square of a number."""
+
+    name: str = "compute_square_tool"
+    description: str = "Compute the square of a number"
+    args_schema: Type[BaseModel] = ComputeSquareInput
+
+    def _run(self, number: str, run_manager: Optional[CallbackManagerForToolRun] = None):
+        """Use the tool."""
+        return float(number) ** 2
+
+    async def _arun(self, number: str, run_manager: Optional[AsyncCallbackManagerForToolRun] = None) -> str:
+        """Use the tool asynchronously."""
+        raise NotImplementedError("Tool does not support async")
+```
+The `name` and `description` of the tool are provided to the LLM when it is invoked and describes what the tool is and what it can be used for.
+
+The `args_schema` describes the inputs the tool expects from the LLM and should be a pydantic class like the `ComputeSquareToolInput`.
+For the agents we've defined, only a single input can be passed to the tool.
+There are other agents that can support structured inputs (multi-input) but the ones in this PoC do not support that.
+
+The execution of the tool happens in the `_run` method.
+The `_run` method should take the input you described in your `args_schema` as a string and return a value that the agent can use to make its next decision.
+Values returned by the tool will be converted to a string before being used by the agent.
+
+### Option 2: Add an API
+If you have a running API (like a websearch API), you can provide it as a tool to the ReAct agent.
+Currently this only supports APIs with JSON inputs and outputs.
+In the `config.yaml` file, replace the "constitution_tool" with the config for your tool.
+You need the following fields in any tools you define (unless it says "Optional").
+
+```yaml
+tools:
+  - name: "constitution_tool" # A unique name for your tool
+    description: "Answers questions about the U.S. Constitution." # A sentence or two describing the purpose of your tool
+    url: "https://www.my.app/v1/completions" # URL to your API endpoint
+    config:
+      method: 'POST' # Type of request (POST, GET, etc)
+      headers: # Headers to send with the request
+        'Content-Type': 'application/json' # Content type must be application/json
+        'Authorization': 'Basic 12345' # Optional: Authorization header (base64 encoded)
+      body: # Key value pairs to send to the endpoint
+        prompt: '{{prompt}}' # The agent's prompt to the tool will be injected anywhere a {{prompt}} is present
+        other-key: 'constant-value' # Optional: Any other key-value pairs to send
+      responseParser: 'json.answer' # Path to the answer to return to the agent
+      responseMetadata: # Optional: Any metadata in the response to capture for use in responseFormat
+        - name: 'sources' # Metadata name
+          loc: 'json.sources' # Metadata path in the response
+      responseFormat: # Formats to provide the response from the tool to the agent, must have agent and json keys
+        agent: '{{response}}' # Response to give to the ReAct agent, the response parsed with the responseParser will be injected anywhere a {{response}} is present
+        json: # Response to give to the Router agent, which will in turn be given to the user
+          - "response" # The response parsed with responseParser
+          - "sources" # Optional: any additional keys to return. Must match one of the name fields in responseMetadata
+    examples: # Optional: Not implemented, provide some example questions your tool can answer
+      - "What is the definition of a citizen in the U.S. Constitution?"
+      - "What article describes the power of the judiciary branch?"
+```

react_agent/__init__.py

@@ -0,0 +1,3 @@
+"""Init."""
+
+__version__ = "0.1.0"

react_agent/agent/__init__.py

@@ -0,0 +1,17 @@
+"""Define agents."""
+
+from langchain.agents import AgentExecutor, create_react_agent
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder, PromptTemplate
+
+from react_agent.constants import REACT_PROMPT
+from react_agent.common.llm import chat_llm, completion_llm
+from react_agent.tools import import_tools
+
+def react_agent():
+    """Create a ReAct agent."""
+    response_format = "agent"
+    tools = import_tools(common_tools_kwargs={"response_format": response_format})
+    prompt = PromptTemplate.from_template(REACT_PROMPT)
+    agent = create_react_agent(completion_llm, tools, prompt)
+    agent_executor = AgentExecutor(name="ReActAgent", agent=agent, tools=tools, handle_parsing_errors=True)
+    return agent_executor

react_agent/api.py

@@ -0,0 +1,58 @@
+from dotenv import load_dotenv
+load_dotenv()
+import os
+import sys
+sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import logging
+from contextlib import asynccontextmanager
+
+import mlflow
+import uvicorn
+from fastapi import FastAPI
+
+from react_agent.agent import react_agent
+from react_agent.apispec import ReActRequest, ReActResponse
+from react_agent.constants import APP_HOST, APP_PORT
+
+logger = logging.getLogger(__name__)
+
+# Start logging
+mlflow.langchain.autolog(log_traces=True)
+
+agents = {}
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """Run startup sequence."""
+    # Add them to the application components
+    agents["react"] = react_agent()
+
+    logger.info("Startup sequence successful")
+    yield
+    agents.clear()
+
+
+app = FastAPI(lifespan=lifespan)
+
+
+@app.post("/react", response_model=ReActResponse)
+async def react(request: ReActRequest):
+    """Interact with the ReAct agent."""
+    agent = agents["react"]
+
+    # Send it to the agent
+    agent_response = agent.invoke({"input": request.prompt, "chat_history": []})
+    answer = agent_response["output"]
+    response = ReActResponse(answer=answer)
+    return response
+
+@app.route("/health")
+def health():
+    """Perform a service health check."""
+    return {"status": "ok"}
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(name)s - %(message)s")
+    uvicorn.run(app, port=APP_PORT, host=APP_HOST)

react_agent/apispec.py

@@ -0,0 +1,17 @@
+from typing import Optional
+
+from pydantic import BaseModel
+
+
+class ReActRequest(BaseModel):
+    """Request for ReAct endpoint."""
+
+    prompt: str
+    tools: list[str] = []
+
+
+class ReActResponse(BaseModel):
+    """Response for ReAct endpoint."""
+
+    answer: str | dict
+    tools_used: list[str] = []

react_agent/common/llm.py

@@ -0,0 +1,50 @@
+from typing import Any, List, Mapping, Optional
+
+import httpx
+from langchain.callbacks.manager import CallbackManagerForLLMRun
+from langchain.llms.base import LLM
+from langchain_core.messages import HumanMessage
+from langchain_openai.chat_models import ChatOpenAI
+
+from react_agent.constants import OPENAI_IGNORE_SSL, OPENAI_MODEL, OPENAI_URI
+
+
+class CustomOpenAI(LLM):
+    """Class to define interaction with the hosted OpenAI instance at a specified URI without SSL verification."""
+
+    base_url: str
+    model: str
+    api_key: str
+    http_client: httpx.Client = None
+    temperature: float = 0.8
+
+    @property
+    def _llm_type(self) -> str:
+        return "custom"
+
+    def _call(
+        self,
+        prompt: str,
+        stop: Optional[List[str]] = None,
+        run_manager: Optional[CallbackManagerForLLMRun] = None,
+        **kwargs: Any,
+    ):
+        # Create the request
+        llm = ChatOpenAI(
+            base_url=self.base_url, model=self.model, api_key=self.api_key, http_client=self.http_client, temperature=self.temperature
+        )
+        request = [HumanMessage(content=prompt)]
+        response = llm.invoke(request, stop=stop, **kwargs).content
+        return response
+
+    @property
+    def _identifying_params(self) -> Mapping[str, Any]:
+        """Get the identifying parameters."""
+        return {"base_url": self.base_url, "model": self.model}
+
+
+verify = False if OPENAI_IGNORE_SSL else True
+http_client = httpx.Client(verify=verify)
+
+chat_llm = ChatOpenAI(base_url=OPENAI_URI, model=OPENAI_MODEL, http_client=http_client, temperature=0, api_key="NONE")
+completion_llm = CustomOpenAI(base_url=OPENAI_URI, model=OPENAI_MODEL, http_client=http_client, temperature=0, api_key="NONE")