@quarry-systems/drift-ai-core

@quarry-systems/drift-ai-core

AI agent core functionality for Drift (Managed Cyclic Graph) - build intelligent workflows with LLM integration.

Overview

Drift AI Core provides utilities and helpers for building AI-powered workflows using managed cyclic graphs. It includes LLM integration helpers, tool calling utilities, schema generation, and runtime type validation.

Installation

npm install @quarry-systems/drift-ai-core @quarry-systems/drift-core @quarry-systems/drift-contracts

Features

• ✅ LLM Integration: Helpers for working with language models
• ✅ Tool Calling: Define and execute tools for LLM function calling
• ✅ Schema Generation: Convert TypeScript types to JSON schemas
• ✅ Runtime Validation: Validate LLM outputs against schemas
• ✅ Streaming Support: Handle streaming LLM responses
• ✅ Message Formatting: Utilities for chat message construction
• ✅ Response Parsing: Extract structured data from LLM responses
• ✅ Error Handling: Robust error handling for AI operations

Quick Start

Basic LLM Integration

import { ManagedCyclicGraph } from '@quarry-systems/drift-core';
import { createLLMNode, formatMessages } from '@quarry-systems/drift-ai-core';

const graph = new ManagedCyclicGraph('ai-workflow')
  .node('chat', {
    label: 'Chat with LLM',
    execute: [async (ctx) => {
      const messages = formatMessages([
        { role: 'system', content: 'You are a helpful assistant.' },
        { role: 'user', content: ctx.data.userQuery }
      ]);
      
      // Use your LLM adapter
      const response = await ctx.services.llm.chat(messages);
      
      return {
        ...ctx,
        data: {
          ...ctx.data,
          response: response.content
        }
      };
    }]
  })
  .build();

Tool Calling

import { defineTool, executeToolCall } from '@quarry-systems/drift-ai-core';

// Define tools
const weatherTool = defineTool({
  name: 'get_weather',
  description: 'Get current weather for a location',
  parameters: {
    type: 'object',
    properties: {
      location: { type: 'string', description: 'City name' },
      units: { type: 'string', enum: ['celsius', 'fahrenheit'] }
    },
    required: ['location']
  },
  handler: async ({ location, units = 'celsius' }) => {
    // Fetch weather data
    return { temp: 22, units, location };
  }
});

// Use in graph
const graph = new ManagedCyclicGraph('tool-calling')
  .node('llm-with-tools', {
    execute: [async (ctx) => {
      const response = await ctx.services.llm.chat(
        formatMessages([{ role: 'user', content: 'What\'s the weather in Paris?' }]),
        { tools: [weatherTool] }
      );
      
      // Execute tool calls
      if (response.toolCalls) {
        const results = await Promise.all(
          response.toolCalls.map(call => executeToolCall(call, [weatherTool]))
        );
        return { ...ctx, data: { ...ctx.data, toolResults: results } };
      }
      
      return ctx;
    }]
  })
  .build();

Schema Generation

import { generateSchema, validateAgainstSchema } from '@quarry-systems/drift-ai-core';

// Define expected output structure
interface UserProfile {
  name: string;
  age: number;
  email: string;
  interests: string[];
}

// Generate JSON schema
const schema = generateSchema<UserProfile>({
  type: 'object',
  properties: {
    name: { type: 'string' },
    age: { type: 'number', minimum: 0 },
    email: { type: 'string', format: 'email' },
    interests: { type: 'array', items: { type: 'string' } }
  },
  required: ['name', 'age', 'email']
});

// Request structured output from LLM
const response = await llm.chat(messages, {
  responseFormat: { type: 'json_schema', schema }
});

// Validate response
const validation = validateAgainstSchema(response.content, schema);
if (validation.valid) {
  const profile: UserProfile = validation.data;
  // Use typed data
}

Core Utilities

Message Formatting

import { formatMessages, addSystemMessage, addUserMessage } from '@quarry-systems/drift-ai-core';

// Format message array
const messages = formatMessages([
  { role: 'system', content: 'You are a helpful assistant.' },
  { role: 'user', content: 'Hello!' },
  { role: 'assistant', content: 'Hi! How can I help?' },
  { role: 'user', content: 'Tell me about graphs.' }
]);

// Helper functions
const withSystem = addSystemMessage(messages, 'Be concise.');
const withUser = addUserMessage(messages, 'What is 2+2?');

Response Parsing

import { parseJSONResponse, extractCodeBlocks } from '@quarry-systems/drift-ai-core';

// Parse JSON from LLM response
const data = parseJSONResponse(response.content);

// Extract code blocks
const codeBlocks = extractCodeBlocks(response.content);
// Returns: [{ language: 'typescript', code: '...' }, ...]

Streaming Helpers

import { handleStreamingResponse } from '@quarry-systems/drift-ai-core';

const stream = await llm.chatStream(messages);

await handleStreamingResponse(stream, {
  onToken: (token) => console.log(token),
  onComplete: (fullText) => console.log('Complete:', fullText),
  onError: (error) => console.error('Error:', error)
});

Tool Execution

import { defineTool, executeToolCall, validateToolCall } from '@quarry-systems/drift-ai-core';

// Define a tool
const calculator = defineTool({
  name: 'calculate',
  description: 'Perform basic math operations',
  parameters: {
    type: 'object',
    properties: {
      operation: { type: 'string', enum: ['add', 'subtract', 'multiply', 'divide'] },
      a: { type: 'number' },
      b: { type: 'number' }
    },
    required: ['operation', 'a', 'b']
  },
  handler: async ({ operation, a, b }) => {
    switch (operation) {
      case 'add': return a + b;
      case 'subtract': return a - b;
      case 'multiply': return a * b;
      case 'divide': return a / b;
      default: throw new Error('Invalid operation');
    }
  }
});

// Validate tool call
const isValid = validateToolCall(toolCall, calculator);

// Execute tool call
const result = await executeToolCall(toolCall, [calculator]);

Advanced Patterns

Multi-Turn Conversations

const graph = new ManagedCyclicGraph('conversation')
  .node('chat', {
    execute: [async (ctx) => {
      const history = ctx.data.messages || [];
      const newMessage = { role: 'user', content: ctx.data.userInput };
      
      const response = await ctx.services.llm.chat([...history, newMessage]);
      
      return {
        ...ctx,
        data: {
          ...ctx.data,
          messages: [
            ...history,
            newMessage,
            { role: 'assistant', content: response.content }
          ]
        }
      };
    }]
  })
  .build();

Agentic Workflows

import { defineTool, executeToolCall } from '@quarry-systems/drift-ai-core';

const graph = new ManagedCyclicGraph('agent')
  .node('think', {
    label: 'Agent Reasoning',
    execute: [async (ctx) => {
      const response = await ctx.services.llm.chat(
        ctx.data.messages,
        { tools: ctx.data.availableTools }
      );
      
      if (response.toolCalls) {
        // Execute tools and continue
        const results = await Promise.all(
          response.toolCalls.map(call => 
            executeToolCall(call, ctx.data.availableTools)
          )
        );
        
        return {
          ...ctx,
          data: {
            ...ctx.data,
            toolResults: results,
            shouldContinue: true
          }
        };
      }
      
      // Final answer
      return {
        ...ctx,
        data: {
          ...ctx.data,
          finalAnswer: response.content,
          shouldContinue: false
        }
      };
    }]
  })
  .edge('think', 'think', (ctx) => ctx.data.shouldContinue === true)
  .edge('think', 'end', (ctx) => ctx.data.shouldContinue === false)
  .build();

Structured Data Extraction

import { generateSchema, validateAgainstSchema } from '@quarry-systems/drift-ai-core';

const extractionSchema = generateSchema({
  type: 'object',
  properties: {
    entities: {
      type: 'array',
      items: {
        type: 'object',
        properties: {
          name: { type: 'string' },
          type: { type: 'string' },
          confidence: { type: 'number', minimum: 0, maximum: 1 }
        }
      }
    },
    sentiment: { type: 'string', enum: ['positive', 'negative', 'neutral'] }
  }
});

const graph = new ManagedCyclicGraph('extraction')
  .node('extract', {
    execute: [async (ctx) => {
      const response = await ctx.services.llm.chat(
        formatMessages([
          { role: 'system', content: 'Extract entities and sentiment from text.' },
          { role: 'user', content: ctx.data.text }
        ]),
        { responseFormat: { type: 'json_schema', schema: extractionSchema } }
      );
      
      const validation = validateAgainstSchema(response.content, extractionSchema);
      
      return {
        ...ctx,
        data: {
          ...ctx.data,
          extracted: validation.valid ? validation.data : null,
          error: validation.valid ? null : validation.error
        }
      };
    }]
  })
  .build();

API Reference

Message Helpers

• formatMessages(messages) - Format message array
• addSystemMessage(messages, content) - Add system message
• addUserMessage(messages, content) - Add user message
• addAssistantMessage(messages, content) - Add assistant message

Tool Helpers

• defineTool(config) - Define a tool with schema and handler
• executeToolCall(call, tools) - Execute a tool call
• validateToolCall(call, tool) - Validate tool call parameters

Schema Helpers

• generateSchema(definition) - Generate JSON schema
• validateAgainstSchema(data, schema) - Validate data against schema

Response Helpers

• parseJSONResponse(text) - Parse JSON from response
• extractCodeBlocks(text) - Extract code blocks
• handleStreamingResponse(stream, handlers) - Handle streaming

Runtime Validation

• isValidToolCall(call) - Check if tool call is valid
• isValidMessage(message) - Check if message is valid
• isValidSchema(schema) - Check if schema is valid

Integration with LLM Adapters

Drift AI Core works with any LLM adapter that implements the LLMAdapter interface from @quarry-systems/drift-contracts:

import type { LLMAdapter } from '@quarry-systems/drift-contracts';

// Your custom adapter
const myLLMAdapter: LLMAdapter = {
  chat: async (messages, options) => {
    // Implementation
    return {
      content: 'Response',
      role: 'assistant',
      finishReason: 'stop'
    };
  },
  
  chatStream: async (messages, options) => {
    // Streaming implementation
  }
};

// Use with Drift
const manager = new Manager(graph, {
  services: {
    llm: { factory: () => myLLMAdapter }
  }
});

Related Packages

• @quarry-systems/drift-core - Core graph engine
• @quarry-systems/drift-contracts - Type definitions
• @quarry-systems/drift-openai - OpenAI adapter

Examples

See the examples directory for complete examples:

• Chat applications
• Tool calling agents
• Data extraction
• Multi-turn conversations

License

Dual-licensed under:

• AGPL-3.0 for open source projects
• Commercial License for proprietary use

See LICENSE.md for details.

For commercial licensing:

• Email: [e-mail]
• Web: https://quarry-systems.com/license

Support

• 📖 Documentation
• 🐛 Issues
• 💬 Discussions