Skip to content

README.md

Latest commit

 

History

History
147 lines (111 loc) · 3.63 KB

README.md

File metadata and controls

147 lines (111 loc) · 3.63 KB

LangChain + Prompt Fusion Integration

This example shows how to integrate Prompt Fusion with LangChain's ReAct agents using the messageModifier callback pattern.

What This Does

Creates a LangChain ReAct agent that:

  • Dynamically fuses base instructions, workspace config, and persona prompts before each LLM call
  • Switches between personas at runtime without restarting the agent
  • Uses semantic weighting to handle conflicting instructions

The fusion happens via messageModifier - a callback that runs before every LLM invocation to inject the fused system prompt.

Prerequisites

  • Node.js 18+
  • OpenAI API key
  • TypeScript knowledge (or use ts-node)

Installation

npm install @langchain/langgraph @langchain/openai @langchain/core zod

Setup

Create a .env file:

OPENAI_API_KEY=your_openai_api_key_here

How to Run

# Using ts-node
npx ts-node agent-with-fusion.ts

# Or using tsx
npx tsx agent-with-fusion.ts

Code Overview

Key Components

1. Prompt Layers:

  • basePrompt - Tool definitions and safety rules
  • brainPrompt - Workspace configuration (Customer Analytics in this example)
  • personaPrompt - Role overlay (Analyst or Researcher, loaded dynamically)

2. messageModifier Function:

return async (messages, config) => {
  const persona = await getPersona(config.thread_id);
  const weights = persona ? { base: 0.2, brain: 0.3, persona: 0.5 }
                          : { base: 0.4, brain: 0.6, persona: 0.0 };
  const fused = fusionEngine.semanticWeightedFusion(...);
  return [new SystemMessage(fused), ...messages];
};

3. Agent Creation:

const agent = createReactAgent({
  llm: model,
  tools: tools,
  messageModifier: createFusionModifier(basePrompt, brainPrompt, getPersona)
});

How It Works

  1. User sends a message
  2. messageModifier runs before LLM call
  3. Checks if a persona is active (via thread_id)
  4. Determines weights: persona active → (0.2/0.3/0.5), no persona → (0.4/0.6/0.0)
  5. Fuses layers with semantic labels
  6. Injects fused prompt as system message
  7. Agent processes with fused context

Customization

Add Your Own Tools

const myTool = tool(
  async ({ input }) => {
    // Your implementation
    return result;
  },
  {
    name: "my_tool",
    description: "What your tool does",
    schema: z.object({
      input: z.string()
    })
  }
);

const tools = [searchTool, myTool];

Change Weight Patterns

// Safety-first (base dominates)
const weights = { base: 0.6, brain: 0.3, persona: 0.1 };

// Context-first (brain dominates)
const weights = { base: 0.2, brain: 0.7, persona: 0.1 };

Add More Personas

const personas = {
  analyst: { /* ... */ },
  researcher: { /* ... */ },
  developer: {
    id: 'developer',
    content: `ROLE: Senior Developer
    SPECIALIZATION: Code review, architecture design...`
  }
};

Troubleshooting

Agent repeats instructions verbatim:

  • Weights might be too balanced. Increase the gap between layers (use 0.1+ differences).

Persona not applying:

  • Check that getPersona() is actually returning content
  • Verify thread_id is being passed in config
  • Log the fused prompt to see what the LLM receives

Tool calls failing:

  • Make sure tool schemas match the actual function parameters
  • Check that tools are properly passed to createReactAgent

Learn More