GitHub - DriftOS/driftos-core: Conversation routing and context management for AI applications. Route messages to semantic branches, extract structured facts with provenance, and assemble optimized LLM context.

driftos-core

LLM-powered semantic conversation routing engine. High-accuracy drift detection using Groq's fast inference for real-time routing decisions.

The Problem

AI applications dump entire conversation history into every LLM call:

Unfocused context → worse responses
Token waste → higher costs
No structure → can't query "what did we decide about X?"

The Solution

driftos-core uses LLM reasoning to detect topic shifts and route messages:

STAY - Same topic, continue in current branch
BRANCH - Topic drift detected, create new branch
ROUTE - Return to a previous topic

Result: Focused context windows. 20 relevant messages instead of 1000.

Why LLM-Based?

Approach	Latency	Cost	Accuracy
LLM-based routing	500-1000ms	~$0.001/call	High
Embedding-based	<200ms	$0	Good

driftos-core uses Groq's fast inference (Llama 3.1) for nuanced understanding of topic shifts, context, and conversational intent. Best for applications where accuracy matters more than latency.

Quick Start

# Clone and install
git clone https://github.com/DriftOS/driftos-core
cd driftos-core
npm install

# Setup database
cp .env.example .env
# Add your GROQ_API_KEY to .env
npm run db:push

# Run
npm run dev

API

Route a Message

POST /api/v1/drift/route
{
  "conversationId": "conv-123",
  "content": "I want to plan a trip to Japan",
  "role": "user"
}

Response:

{
  "action": "BRANCH",
  "branchId": "branch-456",
  "branchTopic": "Japan trip planning",
  "confidence": 0.95,
  "isNewBranch": true,
  "reason": "New topic introduced: planning a trip to Japan"
}

Subsequent Messages

POST /api/v1/drift/route
{
  "conversationId": "conv-123",
  "content": "What's the best time for cherry blossoms?",
  "role": "user"
}

Response:

{
  "action": "STAY",
  "branchId": "branch-456",
  "confidence": 0.92,
  "isNewBranch": false,
  "reason": "Continuing discussion about Japan trip - asking about cherry blossom timing"
}

Topic Shift Detection

POST /api/v1/drift/route
{
  "conversationId": "conv-123",
  "content": "I need to sort out my tax return",
  "role": "user"
}

Response:

{
  "action": "BRANCH",
  "branchId": "branch-789",
  "branchTopic": "Tax return",
  "confidence": 0.98,
  "isNewBranch": true,
  "reason": "Complete topic shift from travel planning to tax matters"
}

Route Back to Previous Topic

POST /api/v1/drift/route
{
  "conversationId": "conv-123",
  "content": "Back to Japan - should I get a JR rail pass?",
  "role": "user"
}

Response:

{
  "action": "ROUTE",
  "branchId": "branch-456",
  "confidence": 0.94,
  "isNewBranch": false,
  "reason": "Returning to previous Japan trip discussion"
}

Get Context for LLM

GET /api/v1/context/{branchId}

Response:

{
  "branchId": "branch-456",
  "branchTopic": "Japan trip planning",
  "messages": [
    { "role": "user", "content": "I want to plan a trip to Japan" },
    { "role": "user", "content": "What's the best time for cherry blossoms?" },
    { "role": "user", "content": "Back to Japan - should I get a JR rail pass?" }
  ],
  "allFacts": [
    {
      "branchTopic": "Japan trip planning",
      "isCurrent": true,
      "facts": [
        { "key": "destination", "value": "Japan", "confidence": 1.0 },
        { "key": "interest", "value": "cherry blossoms", "confidence": 0.9 },
        { "key": "transport_question", "value": "JR rail pass", "confidence": 0.9 }
      ]
    }
  ]
}

List Branches

GET /api/v1/drift/branches/{conversationId}

Extract Facts

POST /api/v1/facts/{branchId}/extract

Endpoints

Method	Endpoint	Description
POST	`/api/v1/drift/route`	Route a message to a branch
GET	`/api/v1/drift/branches/:conversationId`	List all branches
GET	`/api/v1/context/:branchId`	Get optimized LLM context
POST	`/api/v1/facts/:branchId/extract`	Extract facts from branch
GET	`/api/v1/facts/:branchId`	Get existing facts

Configuration

# Required
DATABASE_URL=postgresql://...
GROQ_API_KEY=your-groq-api-key

# LLM Configuration
LLM_MODEL=llama-3.1-8b-instant
LLM_TIMEOUT=5000

# Optional
PORT=3000

How It Works

Analyze - LLM examines message content, conversation history, and existing branches
Decide - Returns STAY, BRANCH, or ROUTE with reasoning
Execute - Creates branch if needed, assigns message, updates state
Extract - When leaving a branch, facts are automatically extracted

LLM Routing Prompt

The routing decision is made by prompting the LLM with:

Current message content
Current branch topic and recent messages
List of other available branches
Instructions to return structured JSON decision

This gives nuanced understanding that embedding similarity can't match - detecting subtle topic shifts, understanding context, and recognizing when someone is returning to a previous discussion.

SDK & MCP

Use with the official SDK:

npm install @driftos/client

import { createDriftClient } from '@driftos/client';

const client = createDriftClient('http://localhost:3000');

const result = await client.route('conv-123', 'Plan my Japan trip');
const context = await client.getContext(result.branchId);
const prompt = await client.buildPrompt(result.branchId, 'You are a travel assistant');

Or use via MCP with Claude Desktop: driftos-mcp-server

Architecture

┌─────────────────────────────────────────────────────────┐
│                     driftos-core                         │
├─────────────────────────────────────────────────────────┤
│  Routes Layer                                            │
│  └── /drift, /context, /facts, /branches                │
├─────────────────────────────────────────────────────────┤
│  Services Layer                                          │
│  ├── DriftService (LLM routing orchestration)           │
│  ├── ContextService (LLM context assembly)              │
│  └── FactsService (LLM-based extraction)                │
├─────────────────────────────────────────────────────────┤
│  Operations Layer                                        │
│  ├── classifyRoute (LLM-based decision)                 │
│  ├── executeRoute (branch/message creation)             │
│  └── extractFacts (structured extraction)               │
├─────────────────────────────────────────────────────────┤
│  Infrastructure                                          │
│  ├── PostgreSQL + Prisma                                │
│  ├── Groq API (Llama 3.1)                               │
│  └── Fastify + TypeScript                               │
└─────────────────────────────────────────────────────────┘

Performance

Routing latency: 500-1000ms (Groq inference)
Accuracy: High (LLM reasoning)
Cost: ~$0.001 per routing decision

When to Use

Use driftos-core when:

Accuracy is more important than latency
You need nuanced understanding of topic shifts
Your conversations have subtle context that embeddings might miss

Use driftos-embed when:

Latency is critical (<200ms)
You want zero LLM API costs for routing
Your topic shifts are clear/obvious

Related Projects

driftos-embed - Embedding-based routing (faster, zero cost)
drift-sdk - TypeScript/JavaScript SDK
driftos-mcp-server - MCP server for Claude Desktop

License

MIT

Patent Pending | driftos.dev