Fire Enrich Documentation

Fire Enrich uses a sequential multi-agent orchestration system to enrich data. Instead of running all searches at once, it executes specialized agents in phases so that each phase can use the context discovered by the previous one.

The Agent Sequence

1. Discovery Agent: Establishes the foundation (company name, website, domain).
2. Company Profile Agent: Uses the verified name to find industry, sub-category, and business model.
3. Financial Intel Agent: Uses company and industry context to search for funding stages, total raised, and investors.
4. Tech Stack Agent: Analyzes HTML, GitHub, and documentation to identify programming languages and infrastructure.
5. General Purpose Agent: Handles custom fields (e.g., CEO, competitors) using the full context gathered in previous steps.

Key Architectural Principles

• Context Building: Each agent adds data that makes subsequent searches more accurate (e.g., industry knowledge helps the Financial agent find the right databases).
• Parallelism within Phases: While agents run sequentially, each agent performs multiple searches in parallel via the Firecrawl API to maximize speed.
• AI Synthesis: A final GPT-4o pass combines all findings, resolves conflicts, and validates the structured data.

Fire Enrich uses a specialized multi-agent architecture to perform data enrichment. Instead of a single generic prompt, it employs five specialized AI agents, each optimized for specific domains:

• Company Research Agent: Extracts company fundamentals (name, description, industry, employee count).
• Fundraising Intelligence Agent: Discovers funding rounds, investors, and valuation data.
• Executive Research Agent: Identifies leadership teams, founders, and key personnel.
• Product & Technology Agent: Uncovers product offerings, tech stack, and competitive landscape.
• Contact Information Agent: Finds emails, phone numbers, and social media profiles.

Agent Mode vs. Traditional Mode

When configuring your enrichment, you can choose between two modes:

• Agent Mode (Recommended): Uses the specialized agents listed above. This provides higher accuracy and better quality extraction for specific fields, though it is slightly slower.
• Traditional Mode: Uses direct GPT-4 extraction. This is faster and uses fewer tokens, making it suitable for simple, non-specialized fields.

To run Fire Enrich locally, follow these steps:

1. Clone the repository:

   git clone <repository-url>
   cd hostedTools

2. Install dependencies using pnpm:

   pnpm install

3. Configure environment variables by creating a .env.local file in the root directory with your API keys:

   FIRECRAWL_API_KEY=your_firecrawl_key
   OPENAI_API_KEY=your_openai_key

4. Start the development server:

   pnpm dev

5. Access the application at http://localhost:3000/fire-enrich.

Prerequisites:

• Node.js 18+
• npm, yarn, or pnpm
• A Firecrawl API key (from firecrawl.dev)
• An OpenAI API key with GPT-4 access (from platform.openai.com)

git clone <repository-url>
cd hostedTools
pnpm install
# Create .env.local
FIRECRAWL_API_KEY=your_firecrawl_key
OPENAI_API_KEY=your_openai_key
pnpm dev

Fire Enrich allows you to select preset fields or define custom fields using natural language.

Preset Fields

You can choose from a variety of standard business intelligence fields, including:

• Company Name
• Industry & Description
• Employee Count
• Revenue
• Headquarters Location
• Social Media Profiles
• Leadership Team

Custom Fields

You can use natural language to instruct the AI to find specific data points. The AI interprets your requirements to generate the necessary extraction logic.

Examples of effective custom fields:

• "Find the CEO's email and LinkedIn"
• "Get their main product pricing"
• "Find recent news mentions"

Tips for High-Quality Extraction

1. Be Specific: Use "CEO name and email" instead of "contact info".
2. Separate Concerns: Request one data type per field.
3. Use Examples: Provide format hints like "Revenue (e.g., $10M ARR)".
4. Leverage Context: Mention your specific use case to help the AI understand the intent.

Fire Enrich is extensible through type-safe Zod schemas. You can add new data points by modifying existing agents or creating entirely new ones.

1. Add a field to an existing agent

Modify the Zod schema in the corresponding agent file (e.g., /lib/agent-architecture/agents/[agent-name].ts).

// Example: Adding a new field to the FundingAgent
const FundingResult = z.object({
  fundingStage: z.string().optional(),
  totalRaised: z.string().optional(),
  lastRoundAmount: z.string().optional(),
  investors: z.array(z.string()).optional(),
  // Add your new field here:
  debtFinancing: z.string().optional(),
});

2. Create a new agent

• Define a new Zod schema.
• Implement the AgentBase interface.
• Update the orchestrator's routing logic to direct relevant fields to your new agent.

Field Routing Logic

The system automatically routes fields to agents based on keywords:

• industry or headquarter $\rightarrow$ Company Profile Agent
• fund or invest $\rightarrow$ Financial Intel Agent
• employee or revenue $\rightarrow$ Metrics Agent
• tech and stack $\rightarrow$ Tech Stack Agent
• Everything else $\rightarrow$ General Purpose Agent

Common Issues and Solutions

Performance & Limits

• Default Processing Rate: 1 row per second (to protect APIs).
• Max Fields: 10 fields per enrichment task.
• Rate Limits: Subject to your Firecrawl and OpenAI plan limits.