Creating a REACT-Style AI Agent with Fireworks AI and LangChain for Data Fetching, SQL Generation, and Conversational Memory

Setting Up the Environment and Tools

This tutorial demonstrates how to build an intelligent, tool-enabled REACT-style agent using Fireworks AI integrated with LangChain. We begin by installing the required packages including langchain, langchain-fireworks, requests, and beautifulsoup4, ensuring the environment is ready for seamless execution.

Importing Necessary Libraries

The project imports essential modules such as HTTP clients (requests, BeautifulSoup), LangChain's agent framework (BaseTool, initialize_agent, AgentType), the Fireworks-powered large language model (ChatFireworks), and utilities for prompt management and conversational memory (LLMChain, PromptTemplate, ConversationBufferMemory). It also uses standard Python modules for secure input (getpass) and environment variable management (os).

API Key Configuration

You are prompted to securely input your Fireworks API key using getpass, which is then stored in the environment variables. This ensures authenticated access to the Fireworks LLM without exposing the key in plain text.

Initializing the Fireworks LLM

A ChatFireworks instance is created using the llama-v3-70b-instruct model with specific parameters such as temperature and maximum tokens. This LLM is ready to process instruction-based prompts immediately.

Example: Sentiment Analysis

A simple sentiment analysis prompt is constructed as a list of messages with system and user roles. The LLM is invoked with this prompt, and it returns a sentiment evaluation of a movie review.

Conversational Memory Implementation

We define a prompt template that maintains conversation history, set up a ConversationBufferMemory to store past interactions, and build an LLMChain combining these elements. This allows the agent to retain context over multiple exchanges, demonstrated by sample inputs and responses.

Custom Tools Definition

Two custom tools are created by subclassing BaseTool:

• FetchURLTool: Fetches the first five paragraphs (up to 500 characters) of text from a specified webpage using HTTP requests and BeautifulSoup parsing.

• GenerateSQLTool: Converts natural language descriptions into commented BigQuery SQL queries by leveraging the Fireworks LLM.

Integrating Tools into a REACT Agent

The custom tools are registered and integrated with the LangChain agent framework using initialize_agent with the ZERO_SHOT_REACT_DESCRIPTION agent type. This setup enables dynamic invocation of tools during conversations.

Running the Agent

An example command directs the agent to fetch content from a Wikipedia page and generate a BigQuery SQL query that counts occurrences of the word "model" in the page text. The generated SQL query is printed as the output.

Extending Capabilities

This integration unlocks a powerful platform for AI applications that go beyond simple text generation. You can extend the agent’s capabilities by adding domain-specific tools, customizing prompts, fine-tuning memory, and exploring advanced features like function-calling and vector-based retrieval to create more dynamic, context-aware assistants.

Additional Resources

Explore the full notebook for detailed implementation. Follow related updates on Twitter, join the Telegram channel and LinkedIn group, and consider participating in the miniCON Virtual Conference on AGENTIC AI for hands-on workshops and certificates.